writeonce/docs/stories/00-status.md

Status board — what is done, what is next

Edges live in 00-dependency-graph.md — mermaid graphs of iteration and feature dependencies; anything with all-green incoming arrows is startable. This board carries the STATES.

The single place to learn where this project stands. Organised in six buckets: stories (the narrative arc), in progress, done, pending, discarded, learnings. The buckets are sections of this board, not folders — a doc stays where it was authored, and only its frontmatter, its banner and this board change.

Three tracks (2026-08-26): language-runtime-database/ — the language and runtime; porch/ — the web framework written in it; databasev2/ — the database beyond RAM. Each numbers its iterations from 1, so a porch 3 is not a language 3; every non-language story carries track: in frontmatter, and moved ones keep was_language_iteration: so a search for the old number still finds them. Track folders are fine; status folders are not.

Status lives in frontmatter, nowhere else (directive 2026-08-26). Every story iteration file sits flat in its track folder and carries status: in its YAML header; the active slice's marker doc sits flat in docs/. No directory anywhere encodes state. This replaces the 2026-08-20/21 convention under which files moved between done/, refine/, hold/ and in-progress/ — those folders are gone. A status change is now a one-line edit, not a move, which is the point: the old scheme broke every relative link in and to a file each time its status changed, and the two link audits (00-link-audit.md) were mostly that.

Every plan and phase doc opens with a > **Status:** banner linking back here; normative contracts (plan/oop-vm/), exploration studies, reference docs and the discarded/learnings registers carry none by design.

Update this board in the same change that finishes work — set the item's status:, record what actually landed here, set the next in-progress item, and record any rejection in discarded.md with its reason.

Two frontmatter axes (2026-08-27), deliberately orthogonal:

• status — where the WORK is: done · in-progress · pending · hold. Rendered here as ✅ · 🔄 · ⬜ · ⏸.
• readiness — whether the DESIGN is settled: ready (brainstorm complete, decisions LOCKED — a spec approved or the forks confirmed) · refine (open forks; cannot be planned yet).

status: refine is retired: it meant both "not started" and "design not settled", so a held iteration with an approved spec was indistinguishable from one nobody had thought about. The startable set is readiness: ready and status: pending.

As of the 2026-08-27 sweep that set has exactly one member: databasev2 4, io_uring group-commit — its four forks were confirmed settled on 2026-08-20 and nothing has started. Across 47 iterations: 19 done, 5 in-progress, 15 pending, 8 hold; 27 ready, 20 refine. So of the 15 pending items only one can be planned without a brainstorm first — which is the number this second axis exists to surface, and it was invisible while one key carried both meanings.

Story frontmatter (iteration/status/chain) is the machine-readable truth behind this board; live Obsidian Dataview views: board-views.md (Kanban = view only, never edits status).

---

▶ NEXT PLAN

Landed 2026-08-30 — keys-resident delta updates DONE, loader refusal lifted

Implemented last time (2026-08-30): the six-task keys-resident delta updates plan's final task — lifting the runtime/src/loader.c refusal of resident: keys and proving update end to end. The refusal (databasev2 2's Outstanding criterion) is now Met: a keys-resident row updates through a WAL delta record, read-modify-append, folded back to a value by wo_wal_fold_row_at on every read, replay and compaction. Proven four ways — the fold itself (earlier tasks), group-commit staging with the id-map re-point deferred to the post-barrier flush, replay/compaction folding delta chains the same way reads do, and this task's oracle test (test_oracle_all_vs_keys_same_update_sequence, runtime/test/test_wal.c) driving the SAME update sequence against a resident: all table and a resident: keys table and asserting byte-identical rows at every step. docs/examples/residency's Product table is genuinely resident: keys now; scripts/residency-accept.sh's gate leg inverted from "the annotation is refused" to "the program runs and place_order's stock decrement survives a restart" (11 checks, 0 failures).

A second bug surfaced auditing the request path before lifting the refusal — the same audit class that caught delete's memory corruption in the prior session. idx_hash, idx_cols_equal and wo_idx_probe (database/src/table.c) read a TEXT column's slot as an engine db_text*, but a keys-resident borrow was handing back VM-decoded wo_str* — a different struct layout, reproduced as a genuine ASan heap-buffer-overflow, not merely wrong values. The same bug was independently present in db.c's GET_FIELD and PROBE arms (inline and request-path), unaudited until now because nothing could reach a keys-resident row through them while the annotation was refused. Fixed at the root: a keys-resident borrow now hands back engine values, exactly wo_row_ptr's contract for resident: all (table.h's own "a row stores NO VM pointer" doctrine) — no index function needed to change, and db.c needed none either. Pinned by test_keys_resident_update_indexed_text, which reproduces the overflow against the pre-fix code; all five pre-existing tests that read a keys-resident Text field directly were auditing the OLD (wrong) contract and are corrected alongside it. test_wal 4746/0 throughout.

What did NOT land, by design — three limitations documented, not fixed: (1) mid-drain stale reads — a request reading a row in the same uncommitted drain as an earlier request's in-flight update to it may see the last durable value, not that write; (2) replay is O(N²) in a row's delta-chain length, since each replayed delta re-folds the whole chain; (3) compaction triggers on byte ratio only, with no per-row delta-count signal, so one hot row (a single popular SKU — this feature's own motivating workload) can grow a long chain without moving the aggregate ratio enough to checkpoint. Item 3 is the sharper finding: the design's decision not to cap chain length rests on compaction bounding it, and for a hot-row workload it does not. Recorded in the story and the example's README.

.dev / reference projects used: none — internal-only, table.c/wal.c/ db.c read directly to audit the request path and trace the representation mismatch.

Dependencies unblocked: none newly technical — databasev2 2's own tasks 6 (the two runtime refusals: no-WO_DATA, the byte budget) and 7 (measure, gate, close out) were already the next items and do not depend on this.

Next steps: databasev2 2 tasks 6/7, as before. database/src/CODE-LOGIC.md is current with the stage-here/commit-in-caller update contract and the engine-representation fix.

Landed 2026-08-29 — databasev2 2 tasks 5c/5d, and a branch consolidation

Implemented last time (2026-08-29): resident: keys storage and every read path. Rows are stored keys-only — the payload dropped after the WAL barrier, the id map holding a log offset instead of a slab slot — and read back through a borrow/release accessor. The scans go through a row iterator that walks the id map for a keys table and the bitmap for a resident one, deliberately, since hash order would reorder every unordered query.

The obligation databasev2 3 left at the compactor is discharged, and it caught more than it predicted. The recorded hazard was that compaction moves records and invalidates stored offsets. True — but the compactor walked the slab bitmap, which a keys-resident row has no bit in, so every such row would have been omitted from the new log outright. Silent data loss, not a bad pointer, and rebuilding offsets would never have caught it. Records are now moved byte-for-byte and re-pointed as they land.

What is NOT done, and why the annotation is still refused: updating a keys-resident row. It lives in the log with no slab slot to mutate, so writing into the borrow's scratch would discard the write silently — the one failure this iteration must not ship. It needs read-modify-append. wo_row_update_field and its slot variant refuse explicitly, and the loader still rejects the annotation, with its message corrected to say so.

Also this session: every branch consolidated to dev and master only. Three could not be replayed and are preserved as annotated tags rather than merged or discarded — archive/cleanup-pre-existing-changes carries the Rust runtime master deleted, and archive/ipc-attach + archive/keypair-auth refactor the same row-API functions db2-keys rewrote. That last one is not a merge conflict but an integration task: iteration 9c transfers ownership of vals on failure, while the keys-resident arm returns early without freeing, so a merge that compiles and passes could still leak or double-free.

porch 1 brainstormed and specced. Phases B and C are superseded before review — both store the response after the handler returns, which cannot detect an in-flight collision at all. Design settled: serialize through a sharded actor pool, persist in a @table. Found and fixed en route: docs/examples/porch/ did not typecheck at all, for want of a use json.

.dev / reference projects used: the PostgreSQL checkpoint study (docs/plan/exploration/postgresql/buffer-and-checkpoint.md) for the compaction shape; the Fiber parity study for porch 1's scope.

Dependencies unblocked: nothing was blocked. databasev2 2's remaining tasks 6 and 7 depend only on iteration 2.

Next steps: databasev2 2 task 6 (the two runtime refusals) and task 7 (measure, gate, close out), then the porch 1 implementation plan.

Landed 2026-08-29 — databasev2 3, WAL checkpoint (the chain's last link)

Implemented last time (2026-08-29): compaction. The log used to grow forever — nothing removed superseded records, so boot replayed all history. It is now rewritten as one record per live row into a temp file and swapped in with rename. Six tasks, brainstormed and spec'd first (spec · plan).

Key findings (measured, not asserted): 2.16× space reclaimed (1 962 358 → 907 094 B), boot 114 → 64 ms, stop-the-world pause 2 651 µs against a stated 50 ms budget. Reading .dev/reference/postgresql was what made the design defensible rather than lazy: Postgres never compacts its WAL, because its records are page deltas and a compacted redo log is not a store — hence heap files, a control file, a redo pointer, a second recovery source and a separate checkpointer process. Ours are full row images, so a compacted log is a complete store, and all of that machinery disappears. What was worth porting is the ordering discipline — publish the switch atomically and last — and one rename provides it.

Learned — two bugs of mine that measurement found, not review: wiring the trigger only into the drain left WO_SHARDS=1 never compacting, its log growing forever (536 KB where multi-shard held 446 KB), because a statement on the owner shard never enters that drain. And the dump was 8× slower than necessary, flushing through the committing path and paying one fdatasync per 256 records for durability that is worthless before the rename — one final barrier took a 2 MB dump from 107 649 µs to 13 212 µs, ~22 MB/s to ~181 MB/s. Separately, the crash battery's first version failed on correct code ~1 run in 3: it acked deletes after committing them, so a kill in between made it demand a row the engine was right to remove. Deletes now announce intent first.

Dependencies unblocked: every link in the concurrency + fiber chain has now landed its planned work — stage 3 → 22 → 24 (absorbing 31 + 34) → 40 → databasev2 4 part A → databasev2 3. Not "complete", precisely: chain 5 stays in-progress because databasev2 4's part B was never done, and its premise was invalidated by part A rather than satisfied. Nothing in the chain is blocked on anything else in it.

Next steps: the honest queue is (1) databasev2 2's outstanding 5c/5d, whose resident: keys half is unimplemented and now carries a recorded obligation — compaction invalidates every WAL offset it stores, so the compactor must rebuild that map; (2) databasev2 4 part B, whose premise was invalidated by part A and which needs re-brainstorming rather than starting; (3) the O(live rows) pause, ~5.5 s at a 1 GB live set, which is the number an incremental checkpoint must be bought against.

.dev/reference used: postgresql — xlog.c (CreateCheckPoint, segment recycling), checkpointer.c (the time-or-volume trigger), and controldata_utils.c, which also corrected a prior exploration doc: Postgres updates its control file in place with a CRC, not by rename.

---

Landed 2026-08-28 — databasev2 4 part A, WAL group commit

Implemented last time (2026-08-28): one durability barrier per drain instead of one per statement. Shard 0 stages every queued write request, holds each reply, commits once when its queue empties, then releases all — so a writer is acknowledged after the barrier that carried its record, which was the intended contract all along and was true before only because every batch had one member. Six tasks, brainstormed and spec'd first (spec · plan).

Key findings (measured, not asserted): ≈2.9× durable write throughput, ≈2.1× lower p50 on a write-concurrent workload, confirmed a second way by the s1-vs-sN split within one build (1467 → 5117 ops/s, mean batch 1.0 → 5.43, peak 57) — 2.9× and 3.5× agreeing. Batching scales with contention: mean batch 1.13 / 1.76 / 5.35 at C = 4 / 16 / 64. The story's premise was wrong: it said "fsync-per-commit" and the engine was fsync-per-statement, committing after every append at all six sites — so part A was closer to deleting calls than adding a mechanism.

Learned — three things the measurement corrected, not the code: (1) /tmp is tmpfs here, where fdatasync is free. The same run reported 195 000 ops/s at p50 1 µs there against 2200 at 7200 µs on ext4. A group-commit measurement taken on a memory filesystem measures nothing; db-bench is right to keep its stores under bench/. (2) No existing leg could exercise the feature — mix writes on one op in ten with C=4, giving 20 writes and mean batch 1.01, so a wmix write-concurrent leg had to be added or the payoff was unevaluable either way. (3) The before-p99 was off the instrument — hist_add clamps at 20000 µs and both before-runs pinned there, so the gain is at least 2.3× and the true old p99 is unknown.

Dependencies unblocked — and one dependency invalidated. WO_T_IO is unreachable from a DB write: a failed stage or barrier now ends the process (exit 74, diagnosed), replacing three behaviours that disagreed — insert un-applied itself while update and delete returned a catchable trap and admitted in their own comments that they left RAM ahead of disk. Part B's premise is invalidated: it was justified by "close the 66× durable gap", but that gap is two problems. Concurrent fan-in was a batching problem and is now ~3× better; a serial writer waiting on one barrier is a latency problem that batching cannot touch and io_uring does not obviously help either. Part B should be re-brainstormed, not started.

Next steps: either re-brainstorm part B against its corrected premise, or take chain 6 (databasev2 3, WAL checkpoint), which now has the replay "before" it lacked. (Superseded 2026-08-29: it landed.) Two debts named rather than hidden: the abort path is not exercised (forcing a real fdatasync failure needs mount privileges), and single-shard concurrent batching needs the inline-path park — the same machinery part B would need.

.dev/reference used: none this slice. The sources were the engine's own code and the Linux fsync-failure semantics that make retrying unsound.

---

Landed 2026-08-27 — iteration 24, chat + actor lifecycle (absorbing 31 + 34)

Implemented last time (2026-08-27): the slice closed and merged to master (ed5334d, fast-forward). T4 monitor + T5 time.after (ids 89/90) had landed on the branch; this session merged master in (adopting the porch rename), finished T8/T9, fixed the gate, found and fixed a runtime bug, and did T10. Iterations 31 and 34 land inside it.

Key findings (measured, not asserted): finishing the gate mattered more than finishing the sample. Making every leg start its own server — instead of the drain leg inheriting the soak's warmed one — exposed that 5 of 16 fresh-server SIGTERM drains left a client at EOF with no close frame and no diagnostic. Traced to shard_main: NEXT_RUNNABLE() already stated the contract ("a WORKER on stop keeps DRAINING … close frames!") but the idle branch reaped and broke, abandoning its inbox. An actor between messages is exactly that idle case. Split out as 40; 20 of 20 clean after. Also measured: the fd check had been core-count dependent — lazy per-shard init takes one io_uring + one eventfd per shard, capped at nproc, so 26 → 44 on a 20-core box read as a leak. 1000 connections left it at 44, which settled it.

Learned: three of the four gate failures were stale build artifacts, not code. A branch switch leaves compiler/_build/ and runtime/build/ holding the other branch's binaries, and a woc emitting .wob v7 against a v6 runtime surfaces only as "no listener" — rebuild both before believing a gate failure. And a gate that reuses another leg's server is not merely untidy: it hid a real bug, and when its own leg failed it orphaned a listener that broke the next run. Example apps now log to /tmp/<app>.log so a developer can tail -F them.

Dependencies unblocked: PUBSUB2 (WebSockets + pub/sub, rejected until this point) is done; the porch ledger's WebSocket rows are ✅ and its cancellation row is unblocked-not-built. Chain position 4 is complete, so the chain's next link is databasev2 4 (io_uring group-commit). Still blocked: CSRF and sessions — iteration 34 shipped HMAC but there is still no RNG, and HMAC authenticates a token without being able to mint one, which is 39's leading item.

Next steps: databasev2 4, or databasev2 2's outstanding 5c/5d. One debt is named rather than hidden: iteration 40's guarantee is proven only by the chat gate — nothing in runtime/test/ drives wo_engine_start/wo_engine_stop and no corpus fixture can trigger a stop, so pinning it lower needs new multithreaded test infrastructure.

.dev/reference used: none this slice. The sources were RFC 6455, RFC 3174/4231 for the digest vectors, and the kernel's own interfaces for the drain.

Landed 2026-08-27 — databasev2 1, the RAM ceiling measured

Implemented last time (2026-08-27): databasev2 1 refined (three forks settled) and implemented. A text-heavy Wide reference shape beside the Int-only Item; growth N int|text in the db-bench sample, reading its OWN /proc/self/status RSS at each decile because the driver's 250 ms poll misses the value at a boundary; growth-verify, which asserts the survivor of a crash is a contiguous intact prefix; and two harness legs — four footprint legs under a rootless cgroup v2 cap, a ceiling leg that deliberately dies at the cap and then replays, a randread leg that reads an oversized table randomly, and a replay leg that times boot against history length. 148 checks, 0 failures.

Key findings (measured, not asserted): per-row footprint is 96.5–100 B Int-only and 320.6–324 B text-heavy — 3.3×, not the "order of magnitude" three docs asserted. Read as the median of per-decile marginals, never a two-point slope: index doublings make a two-point read swing 2× (96 vs 205 B/row for one shape). Two predictions in the iteration's own premise were wrong. The ceiling is not a catchable WO_T_OOM for table storage — it is SIGKILL, signal 9, because vm.overcommit_memory = 0 lets malloc succeed and the kernel kills on page touch, so the checked path never runs (the VM arena is the opposite: WO_HEAP_MB is checked and traps). And swap is not "latency collapse": 900 000 rows inside a 64 MiB cap with swap finished in 148 s against 150 s uncapped — ~1%, on a real disk swap file with no zram. Also measured: ack-after-fsync holds through an OOM kill — ~40 000 rows came back as an intact prefix, no holes, not read as corruption. And the pattern the swap leg was missing: random reads over an oversized table collapse 273×. Plus replay: ≈5.5 µs per WAL record, and 1.9× the boot cost for an identical live dataset once each row has been updated once (110 → 211 ms for the same 20 000 rows) — boot replays history, not data.

Learned: an append-mostly workload never re-touches its cold pages, so swap costs it nothing — and the opposite pattern was then measured on the same day. The randread leg reads randomly across a table larger than the cap, both legs walking the SAME Weyl key order so residency is the only variable: 273× throughput collapse (1 851 166 → 6 771 reads/s), p99 1 µs → 487 µs, all 20 000 reads resolving in both. So the two access patterns sit ~270× apart under identical memory pressure, and departure is a step, not a curve — which is why p99_departure_decile finds nothing: there is no knee to find. The RAM ceiling therefore has two shapes and neither announces itself: without swap the process vanishes on signal 9, with swap it keeps returning 0 while serving from disk. That is the argument for a budget that fires at a declared threshold instead of at exhaustion.

Dependencies unblocked — one, by removing it: iteration 2's resident-footprint budget default was to be derived from "swap onset". There is no onset. Swap-off jumps straight from working to SIGKILL; swap-on shows no degradation to detect. Iteration 2 must pick its budget on other grounds rather than wait on a number this slice cannot produce. Iteration 3's replay baseline is still NOT delivered — bench/baseline.json times no replay.

Next steps: iteration 2's 5c/5d. Its task 7 gained a criterion from this: resident: keys must measure its OWN read path rather than inherit 273×. That number bounds demand-paged anonymous memory through swap (4 KiB per fault, no readahead); pread through the page cache should beat it, and the entire value of resident: keys rests on how much — if it is not materially better than swapping, the design buys nothing the kernel was not already doing.

Iteration 3 now HAS its before, and a correction: it planned to use "22's aged-store replay numbers", which never existed — 22 proved restart correctness and never timed it. Also recorded there, before it could be rediscovered late: compaction invalidates every resident: keys offset, since it rewrites the log and moves every record. Not stale-but-readable — an arbitrary byte in a rewritten file. So compaction cannot be a pure file operation that ignores in-memory table state.

.dev/reference used: none. Sources were the kernel's own interfaces — cgroup v2 memory.max/memory.swap.max, /proc/self/status, /proc/swaps and vm.overcommit_memory.

---

Landed 2026-08-25 — packaging + release pipeline (off-chain, no story)

Implemented last time (2026-08-25): the toolchain became installable by a stranger. VERSION as the single source (0.1.0, asserted against both binaries by scripts/mkdist.sh), just dist producing writeonce-<ver>-linux-amd64.tar.gz + .sha256 with scripts/install-readme.tmpl.md inside it, just install-accept proving a from-scratch project builds against the extracted tarball's own binaries, and .github/workflows/release.yml publishing on a v* tag push. Runbook: guides/releasing.md.

Key findings (measured, not asserted): the build host's glibc caps what the shipped binaries can import, and that cap becomes every user's floor — so runs-on is ubuntu-22.04 (2.35) deliberately, not ubuntu-latest; built on this dev machine the binaries need GLIBC_2.38, which would silently exclude Ubuntu 22.04, Debian 12 and RHEL 9. ocaml/setup-ocaml@v3 gives a compiler and opam but not dune, and pinning dune.3.14.0 is a downgrade the solver refuses — take whatever it provides, since any dune ≥ 3.14 satisfies (lang dune 3.14).

Learned: the asset filename is load-bearing. /install links one exact URL, so the workflow asserts tag = VERSION = asset name and fails rather than publishing a download button that 404s. A measurement that only prints is not a gate — the glibc floor is printed from the artefact about to ship, so the claim on the page can be checked against a build log instead of trusted.

Dependencies unblocked: nothing in the chain; this is the distribution seam. It does make docs/examples/site's /install page truthful, which iteration 37's site restructure had left pointing at an asset nobody had built.

Next steps: the live slice is iteration 24, untouched by this. CI is release-only — no workflow runs the gates per change, which remains the open half of iteration 30 (observability, CI, fuzz — still no story file).

.dev/reference used: none — GitHub Actions' own docs and the runner images' glibc versions were the only sources.

---

Landed 2026-08-25 — iteration 37, wo-html components (off-chain)

Implemented last time (2026-08-25): iteration 37 CLOSED, both halves. The grammar half (2026-08-24) added the backtick raw text literal — content verbatim, common margin removed at lex time, ${ } raw and {{ }} compiling to a call on the esc in scope — plus WO-E004/WO-E005. The library half (2026-08-25) added Component, render_all and Layout to wo-html, moved ok_html into framework/http beside ok_text/ok_json, and migrated BOTH HTML samples onto the component layer.

Key findings (measured, not asserted): multi Component holds a heterogeneous list DIRECTLY — no wrapper record — so the framework's Mw/Aw shape is a local choice, not a language requirement; that is what made page components able to own their children. The whole escaping desugar needed zero compiler knowledge of HTML: {{ e }} is a Call on an ordinary in-scope esc, so typecheck, ownership, codegen, the .wob format and the VM were all untouched. {{ }} was proven byte-identical to the hand-written esc() calls it replaced, hostile input (< > & ") included, across all eight migrated builders.

Learned: an interface that nothing consumes as a TYPE is decoration — Component only started earning its place once render_all and the page components held multi Component. And the shop template DOES build and run — an earlier note in this repo had that wrong, and wrong again about why: gap #1 (pub + @table) constrains neither the build NOR the layout. A class crosses module lines without export; only a free fn is module-scoped (WO-E210).

Dependencies unblocked: shop README gap #3 ("no multi-line expression or literal") is closed. Separate .html templates, if ever wanted, now have exactly one honest shape — a COMPILE-TIME include feeding the raw-literal machinery; a per-request file read is the already-rejected engine.

Next steps: the concurrency chain below is untouched by this and remains the live queue.

.dev/reference used: none this slice (Angular's component format was studied from its public docs during the 2026-08-23 story write-up; no reference project was consulted for the implementation).

---

The concurrency + fiber chain — ✅ stage 3 → ✅ 22 → 🔄 24 (absorbing 31 + 34) → 23 → 32. The chain's original order put 31 before 24; the 2026-08-23 directive absorbed 31 INTO 24, and 34 resolved with it, so those three are one slice. Iteration 24 is nine of ten tasks landed and MERGED TO MASTER on 2026-08-27 (fast-forward, ed5334d): T1 crypto, T2 bounded mailboxes, T3 call/reply, T4 monitor + T5 time.after (ids 89/90 — the reserved holes are now filled), T6 ws upgrade, T7 frame codec, T8 chat sample, T9 the chat gate. Verified on master: chat 11 checks 0 failures at the full 1000-client soak, runtime battery 36 suites 0 fail, compiler 556 checks 0 fail, corpus 119 checks 0 fail. Only T10 closeout remains — which is what still holds stories 24/31/34 open. Finishing T9 exposed and fixed a real runtime bug, split out as 40. Its running state is the marker doc (the marker doc, deleted at closeout per the convention), which is the file to read for what is done and what is next; stories 31 and 34 keep status: refine until 24's T10 closeout sets all three to status: done together.

Implemented last time (2026-08-21): iteration 22 landed — the measurement backbone exists and every performance claim is now sourced. docs/examples/db-bench + scripts/db-bench.py + bench/baseline.json (74 metrics, tolerance-tuned by a two-run repeatability check) + just db-bench/db-bench-quick; time.ticks (µs monotonic clock, builtin 84) as the one runtime addition. Restart proof + 3× kill -9 battery per shard count all green; the gate bites (doctored results fail on exactly the doctored metric).

Landed 2026-08-22 — the read-path index slice (born from the postgres study + 22's numbers, commit 6a306a7): engine wo_idx_probe (bucket lookup, scan-identical verify) + emitter index selection (where var.col == key lowers to DB_PROBE; guards stay the arbiter). Reads 1.3k → 1.3M ops/s, p50 600µs → 1µs (~×850); mixread 21 → ~1.9k ops/s multi-shard. Baseline refreshed; tolerance policy moved into the driver (refresh-proof); gate proven to bite on both classes. Pinned by corpus query-index-probe + a wo_idx_probe unit suite.

Key findings (measured, not asserted): durable seed ≈4.5k inserts/s vs ram ≈297k/s — the 66× fsync gap IS iteration 23's case; point lookups WERE O(table) (fixed 2026-08-22, above); mixread was 1,280 ops/s single-shard vs 21 ops/s multi-shard — the DB-actor price under O(table) probes and owner serialization; msgrate 13.4M msgs/s same-heap vs 2.45M cross-shard — deviation 4's mutex-inbox number (rings stay unearned until this is the bottleneck). Standing bug found: hand-built multi <TableClass> SEGVs on drop (elements classed OWNED; refs are scalar ids).

Learned: benchmark tolerances must be per-class — mix* spreads 50% run-to-run (scheduling), read/query jitter ~25%, seed/write/msgrate hold at 15%; a RAM store dies with its process, so throughput modes share one run (all); WO_DATA on tmpfs makes fsync free — durable numbers need a real disk.

Dependencies unblocked: 23 (has its fsync baseline to beat), 31 (has the mutex-inbox number), 32 (has the restart/replay timing machinery), and every future optimization (the gate that catches regressions is live).

Next steps: finish 24 (T4 monitor id 89, T5 time.after id 90 — both still literal holes in wob.h's builtin enum; then T8 the chat sample, T9 its gate, T10 closeout setting 24/31/34 to status: done) → 23 (io_uring group-commit — target: close the 4.5k→297k durable gap) → 32 (WAL checkpoint). Held tail resumes on its own precedence notes.

(Superseded 2026-08-28: 24 landed, and 23's part A landed with it — "close the 4.5k→297k durable gap" turned out to be the wrong target; see the databasev2 4 row.)

.dev/reference used: none this slice (the LW_SOAK discipline and linkcheck.py precedent came from in-repo scripts).

---

Landed 2026-08-20 — framework v1 (the previous NEXT PLAN)

Framework v1 — a polished micro-framework (routing, middleware, Req/Resp), nothing MVC-scale. Directive 2026-08-20: iteration 17 (library kind + internal/) is parked — spec + plan approved and ready on branch library-internal — and the framework itself is the work. The v1-polish slice landed the same day (branch framework-v1): registration helpers get/post/put/delete_ (the take-Handler shape, probe-proven), 405 + Allow on wrong-method hits, HEAD served as GET with the body suppressed, a Logging middleware, set_header; just web-app grew to 16/0. En route it exposed and fixed a real emitter bug: a Text-typed single-segment interpolation of a place ("${r.method}", r a loop borrow) crossed let/assignment boundaries uncopied — aliased the field, crashed the release build; copy_place_text now sees through Interp exactly as drop_fresh_text does, pinned by tests/corpus/run/interp-borrowed-field.

Auth-in-core landed 2026-08-20 (same branch): http/auth.wo — header parsing, pure-.wo base64, constant-time ct_eq, req.principal as the blessed principal slot (Middleware.before takes mut req), BearerAuth + BasicAuth middlewares; policy stays app-side. Probe matrix 26/26 ASan-clean; web-app dogfoods BearerAuth. The framework README carries the core checklist (✅ / candidate / parked-by-design rows).

Form-encoded bodies landed 2026-08-20 (same branch): media_type(req)

• form_values(req) (nil on any other content-type; '+'/%XX decoded); CreateProduct accepts form OR JSON into one insert path.

Multipart landed 2026-08-20 (same branch): http/multipart.wo — RFC 7578 fields + file parts, strict malformed-is-nil, part_named; whole-body within BODY_MAX (streaming parks behind 8/11). CreateProduct takes multipart/form/JSON; just web-app 21/0. Surfaced + fixed the RETURN flavor of the interp-of-borrowed-place emitter bug (emit_return now sees through Interp; same corpus pin). Body-parsing hooks: all three ✅.

Scope split (2026-08-20): the surface above plus the remaining transport/ routing/security gaps is framework v1, tracked item-by-item in the framework README's status ledger (✅/🔶/⬜/⏸/🔧 per feature — timeouts and Unix sockets need net runtime seams, crypto hashes need C builtins since the language has no bitwise operators, streaming/cancellation park behind 8/11). The memory-rich features are framework v2 = iteration 18 (⏸ HELD 2026-08-21 with spec approved + plan authored intact): TTL cache, @table flags, durable job queue with drain-on-request, transaction { } over the WAL's staged batch. The pending order is now the concurrency chain (see Pending below). Edges: 00-dependency-graph.md.

---

Landed 2026-08-15 — the executable milestone (the previous NEXT PLAN)

"Make log-watcher executable" — the difference between "it runs" and "you can leave it running". Every item came from a measurement on the sample itself, and all six landed:

1. The ownership pass does not know what the stdlib returns — done 2026-08-14. The root cause was deeper than the table: Text was classified Copy, so no Text local was ever dropped. Text is now an owned heap value that is copied at every ownership boundary (container, field, return, binding, loop cursor), the ownership pass reads the stdlib, builtin and static tables, and fs.read_all/net.read no longer mis-size a short read's buffer. Measured: run 1 051 040 B → 2 112 B, watch 128 B → 64 B; what remains is items 2 and 3 below, by stack.
2. A projected temporary is never dropped — done 2026-08-14. The projection was one of six shapes with no owner: a call result compared against nil, an argument the callee only borrows, a container read's copy, a loop's iterable, a projected record, and any of those escaped by a return from inside the statement that built them. Measured: run 2 112 B → 64 B and flat from 8 s to 20 s, the full MCP mix 21 312 B / 63 → 64 B / 1, every handler flat from 2 to 6 requests. The 64 bytes left are item 3, on every path.
3. The runtime leaks its own argv container — done 2026-08-14. The entry only borrows its arguments, so main.c releases the container it built, after the entry returns and after a trap alike. All three modes now report ZERO leaks under ASan — watch, run, and the full MCP mix — which is the clean baseline item 6's soak needs to read against.
4. A stopping program does not stop — done 2026-08-14. A blocking call that parks (net.accept, socket read/write, time.sleep, a child wait) now ends the program when it is interrupted with the stop flag set, instead of restarting the syscall. A stop is not a trap: try cannot swallow it, and the stack unwinds through the same drop machinery, so the exit is clean and leak-free in every mode. It also uncovered a real double-free: an assignment of a Text place was a move, not a copy, so api_key = j.mcp.apiKey aliased the record — let copied, assignment now does too. just log-watcher is 7 checks; the seventh is the stop.
5. The MCP server never closes an accepted connection — done 2026-08-15. net.close on every path out of a serve iteration (and the listener on stop). Measured: 4 → 4 descriptors across 200 requests, was one leaked per request.
6. Nothing soaks — done 2026-08-15. LW_SOAK=<seconds> drives all three modes under load and fails on resident growth past 256 KiB or any descriptor growth. The soak immediately caught what every seconds-long check missed: ~1.6 MiB/min of in-arena leaks the ASan report cannot see (the arena is one allocation to LeakSanitizer). Five bugs fell out: json decode's worst-case string sizing (free lists poisoned by relabeled lengths), != never dropping fresh operands, Int interpolation segments mistaken for borrows, json.encode(Ctor{...})'s unowned argument, and discarded statement results (pop(lines);). After: release soak 30 s per mode — watch 0, run 0, mcp +20 KiB, descriptors flat; ASan build flat at 14 600 KiB across 601 686 requests in 90 s once past its ~1200-request quarantine warm-up.

Plan: plan/compiler/2026-08-14-logwatcher-executable.md · Story slice: docs/stories/language-runtime-database/07-logwatcher-proof.md

Deferred by name, with the measurement that says so:

• Iteration 5's strictness half — ?T forced handling landed 2026-08-18 (WO-E211/212/213 + local narrowing; the samples were updated to the bind-then-narrow idiom and stay green), reject rows landed 2026-08-18 (WO-E105 doctrine diagnostics — super.f() used to compile clean). Still open: pub(read) write enforcement, using, #if. Plan 8 stays open.
• Everything @gc: iteration 7b, set's @gc retention gap, iteration 4's gc/held-cycle leak. The sample declares no @gc class — 35 classes, none with the gc flag, 0 RC_INC/RC_DEC against 78 DROPs — so none of it can affect this workload.
• Iterations 8–12 (shard-actor runtime, database engine, @table/query, HTTP layer, fibers, blue-green): unchanged, and unblocked by this plan.

The language track's first goal — iterations 3 → 4 → 5 → 6 → 7, compile and run log-watcher — is met; the database engine (9/9b), deps (15), and the web framework (16) landed on top of it. The goal is now the framework as a polished micro-framework (17 parked; see the NEXT PLAN above and "Implementation order" under Pending). (The prior Rust wo runtime was removed from the repo 2026-08-18 — see discarded.md.)

---

Stories

docs/stories/language-runtime-database/ — one language, one runtime, one database, one binary. Twelve iterations, each an unsplittable slice with Given/When/Then acceptance and a pointer to the plan that sequences its tasks. Read one, approve, then the next starts.

#	Iteration	State
1	Principles doc	✅
2	VM core (wovm)	✅
3	Compiler front (woc)	✅ (known gaps below)
4	Single binary end-to-end	✅ (known gaps below)
5	Language surface	🔄 grammar done; ?T forced handling ✅ + reject rows ✅ + WO-E205 ✅ (2026-08-18); pub(read)/using/#if still ⏸
6	Program mode + stdlib	✅ (the surface log-watcher uses)
7	log-watcher proof	✅ landed 2026-08-15 — executable, not merely compilable: zero ASan leaks in all three modes, SIGTERM ends parked syscalls, fds flat, LW_SOAK gate; just log-watcher 7/0
7b	Inferred GC + mark-sweep	✅ landed 2026-08-18 — @gc gone (WO-E104), GC-ness inferred, RC replaced by incremental mark-sweep, .wob v4; supersedes iteration 2's RC memory model
8	Shard-actor runtime	✅ landed 2026-08-21 — the arc complete: stages 1+2 (fibers/budget/actors/io_uring plane, shards, envelopes, WO-E222) + stage 3's transparent DB actor (just db-actor 8/0, ASan/TSan clean, WAL replay pair)
9	Database engine	🔄 engine complete (storage/WAL/indexes/insert-update-delete); reads land with 9b
9b	@table, relations, query	🔄 query surface + relations + FK done (branch query-surface); group-by parked
19	Float + Bytes	✅ landed 2026-08-20 — .wob v5: Float constant tag, field kinds 6/7, opcodes 34-41 (IEEE-quiet f64), builtins 70-83. Full stack: literals, arithmetic, @table column, WAL bit-exact replay, json fractions in / shortest-round-trip out, ?Float reserved-NaN nil, total-order index (NaN last, -0.0 == +0.0), Bytes + base64. No implicit Int/Float mixing (WO-E201); float/trunc are the only bridges. Proof: web-app price is a real Float ({"price":9.99}), just web-app 23/0; corpus 103/0
11	Fibers	✅ landed 2026-08-21 with the arc (just fibers 10/0); fs-park re-scoped out of v1, disclosed in the story
22	Durability, throughput, scale	✅ landed 2026-08-21 — db-bench + baseline.json (74 metrics) + restart/kill -9 proofs both shard counts; durable 4.5k vs ram 297k inserts/s, reads O(table), msgrate 13.4M/2.45M
31	Actor lifecycle	✅ LANDED 2026-08-27 inside 24 (directive 2026-08-23). All four mechanisms: call/reply with a typed scalar reply (WO_B_CALL = 88, WO-E226), bounded mailboxes (WO_MAILBOX, cap 1024, catchable WO_T_ACTOR), actor death that traps callers instead of hanging them, monitor (89) and time.after (90) — the reserved holes in wob.h are filled. A fifth mechanism it did not anticipate came out of proving the gate: the shutdown drain guarantee, 40. Supervision trees stay out of v1
24	chat: WebSocket workload	✅ LANDED 2026-08-27 (absorbing 31 + 34) — all ten tasks; merged to master ed5334d. just chat 11 checks, 0 failures at the full 1000-client soak: handshake, functional matrix on both WO_IO backends and on one shard, the soak, the fd invariant, the SIGTERM drain, WO_MAILBOX=8 backpressure, ASan clean. Finishing its gate found a real runtime bug, split out as 40
23	io_uring group-commit	✅ part A LANDED 2026-08-28 — group commit, one barrier per drain instead of one per statement (the engine was fsync-per-STATEMENT, not per commit; the story's premise was wrong). Shard 0 holds each reply, commits once when its queue empties, releases all — so a writer is acked after the barrier carrying ITS record. ≈2.9× durable write throughput, ≈2.1× lower p50, two measurement methods agreeing (2.9× controlled, 3.5× s1-vs-sN); mean batch 5.43, peak 57. A durability failure is now fatal (exit 74), not a catchable WO_T_IO — replacing three behaviours that disagreed, two of which admitted leaving RAM ahead of disk. What it did NOT do: durable.sN.mixwrite 480→492 (unchanged — that workload does 20 writes at C=4, mean batch 1.01) and seed unchanged (serial writers have nothing to batch with). This row used to say "close the 66× gap"; that target was mis-stated — the gap is two problems and part A fixes only the concurrent one. ⬜ part B (io_uring) needs re-brainstorming, not starting on the old premise
32	WAL checkpoint	✅ LANDED 2026-08-29 — the chain's last link. Compaction rewrites the log as one record per live row and swaps it in with rename, so recovery is completely unchanged and crash safety comes from the filesystem rather than from code. 2.16× space reclaimed (1 962 358 → 907 094 B), boot 114 → 64 ms, stop-the-world pause 2 651 µs against a stated 50 ms budget. Read .dev/reference/postgresql for it: PG never compacts its WAL — its records are page deltas, so it needs heap files, a control file, a redo pointer and a separate process. Ours are full row images, so a compacted log IS a store, which deletes all of that. kill -9 during compaction: 40 rounds/run, 10 clean runs, and mutation-proven — against in-place rewrite instead of rename the battery fails every time. Outstanding: the resident: keys offset map (compaction moves every record; the obligation is recorded at the compactor) and the O(live rows) pause, ~5.5 s at 1 GB, which is what an incremental design must be bought against
33	Single-file store	⬜ off-chain, small — WO_DATA=<path>.db file form; driver-only (story written 2026-08-22)
34	Crypto builtins	🔄 code landed as 24's T1 (d14fa9f): sha1/sha256/hmac_sha256, ids 85–87 in wob.h, runtime/src/crypto.c, RFC/FIPS vectors 18/0, corpus pin. The 24 gate that once needed it is cleared. Frontmatter keeps status: refine only until 24's T10 closeout sets it to done
38	Content platform capabilities	⬜ off-chain, needs a spec — the two capability families no iteration owns, confirmed against runtime/src/wob.h: fs mutation verbs (six fs builtins, ids 40–45; append creates-if-absent, so nothing is ever replaced, truncated, deleted or renamed) and net.connect (ids 51–55 + 91–95, no connect, and no connect() anywhere in runtime/src/ — so no OIDC/SMTP/object-store/webhook/federation). Driven by a docs/examples/vault content-collaboration workload, in 28's mould. New builtins from 96 (89/90 reserved for 31); no .wob bump (WOB_VERSION 6u, last moved by 36). Story written 2026-08-26 from the "can it build a Nextcloud?" ask
39	Web framework parity	⬜ off-chain, needs a spec — from the Fiber v3.5.0 study (all 32 of its middleware read against porch; nine already have a counterpart). Leads with a random-bytes builtin: the framework ledger claimed CSRF/sessions were unblocked by iteration 34's HMAC, but HMAC authenticates a token and cannot mint one — there is no RNG anywhere in the runtime. Then cookies (absent both ways; Resp.headers being a map cannot carry two Set-Cookie lines), then limiter/idempotency (cheapest wins — @table + time.ticks, nothing new), sessions, CSRF, and the routing/response sugar. Streaming/SSE/compression, @derive binding, TTL cache, proxy and metrics all excluded with owners named
40	Shutdown drain guarantee	✅ LANDED 2026-08-27 — chain 3, with 31; split out of 24. One rule: a message sent before the stop flag is observed must be delivered and run before the engine stops. Found by measurement, not review: making the chat gate's drain leg start its OWN (cold) server exposed that 5 of 16 fresh-server SIGTERM drains left a WebSocket client at EOF with no close frame and no diagnostic. Traced to shard_main — NEXT_RUNNABLE() already stated the contract ("a WORKER on stop keeps DRAINING … close frames!") but the IDLE branch reaped and broke, abandoning its inbox for teardown to free. An actor between messages is exactly that idle case, which is why a WARM soak server hid it for so long. Fix is one branch honouring the primary's drain window, yielding on an empty poll. 20 of 20 clean after; just chat 11 checks 0 failures at the full 1000-client soak (which also settled the fd question: 1000 connections left the count at 44); runtime battery 36 suites 0 fail, compiler 556 checks 0 fail. Ruled out: a bigger spin (a 1 s wall-clock deadline still failed 2 of 12) and spawn-during-shutdown. Outstanding: a pin below the gate — nothing in runtime/test/ drives the engine start/stop and no corpus fixture can trigger a stop
37	wo-html components	✅ off-chain — LANDED 2026-08-25. Raw text literal (backtick, margin stripped at lex time, {{ }} auto-escapes) + the component layer: Component/render_all/Layout in wo-html, ok_html moved into the framework, site and shop both migrated
35	net runtime seams	⬜ off-chain — fd deadlines on the park plane, Unix sockets, peer address; owns the ledger's three 🔧 rows (story written 2026-08-22)
25	HTTP service layer	⏸ hold (2026-08-21) — story file removed; the plan doc remains
26	Blue-green deploy	⏸ hold (2026-08-21)
28	skillhost host workload	⏸ hold (2026-08-21); gaps recorded (branch query-grammar found skillhost needs no new query grammar)
29	Compile-time metaprogramming	⏸ hold (2026-08-21)
15	deps: wo.toml [deps]	✅ landed 2026-08-18 (branch web-framework): [deps] inline tables, git-binary fetch, wo.lock pinning, offline-when-locked, --update-deps, WO-E106/E107; just deps-accept 8/0
16	web framework	✅ landed 2026-08-19 — writeonce-framework (HTTP/1.1 + router + Handler/Middleware) consumed by web-app through [deps]; h2c parked (§C) behind 8/23/11. v1 polish landed 2026-08-20 (branch framework-v1): get/post/put/delete_ helpers, 405+Allow, HEAD, Logging middleware, set_header; just web-app 16/0; fixed the interp-borrowed-field emitter crash en route. Auth-in-core landed 2026-08-20: http/auth.wo (Bearer/Basic, ct_eq, req.principal), web-app dogfoods BearerAuth, gate 17/0
17	library projects + internal/	✅ landed 2026-08-20 — kind = "library" in wo.toml (default program, so every existing manifest is byte-identical; unknown value = WO-E109 exit 2); woc <dir> on a library runs the FULL pipeline entry-less and writes nothing, retiring iteration 16's --emit workaround; the no-entry build error names the kind; lib+bin dual works. Go's internal/ rule as WO-E108 at the consumer's own use, dep-boundary-only — the library imports its own interior freely. Framework reorganized: internal/{parse,serve}.wo behind the line, http/form.wo split out to keep media_type/form_values public. Driver-only change; VM/.wob/GC untouched. just web-app 26/0 (3 new checks), every standing gate unchanged
18	framework v2: memory-rich features	⏸ hold (2026-08-21); spec approved + plan authored, both held intact (spec, plan): TTL cache + @table flags + durable job queue (drain-on-request) + transaction { } over the WAL's staged batch; pub/sub rejection expired with the arc (8/11 landed 2026-08-21) — revisit on unhold

---

In progress

Track	Item	Where
Language	🔄 iteration 36 — operator parity: not, bitwise & | ^ << >>, hex/binary/_ literals, compound assigns — CODE LANDED 2026-08-22 (branch operator-parity, .wob v6, all gates green; reference project .dev/reference/go drove the design). Awaiting the developer's MANUAL pass on docs/examples/operators/ (no test fixtures by directive); unblocks story 34's pure-.wo HMAC question	plan
Language	the framework v1-polish slice landed 2026-08-20 (branch framework-v1, awaiting merge); next per the order: brainstorm 20/21's forks	order
Runtime	✅ iteration 35 landed 2026-08-23 (branch framework-v1b, with framework v1 slice 2 + the serving slice): net deadlines/unix/peer (ids 91–95), fiber pooling, serve_conn + web-app fiber-per-connection — web-app gate 41/0, both WO_IO backends	design

No slice is active. Iteration 24 landed 2026-08-27 and its marker doc was deleted per the convention. Everything pending is the concurrency chain (see Pending below) — the chain's next link is databasev2 4 (chain 5, the io_uring group-commit write path, was_language_iteration: 23), which now has iteration 22's fsync-per-commit numbers in hand, plus databasev2 1's finding that the write path is not where memory pressure bites (appending under a cap costs ~1%, random reads 273×). The held tail is every story whose frontmatter reads status: hold.

Landed 2026-08-14 — the compile-and-run milestone

One session, driven end to end by compiling docs/examples/log-watcher and watching its diagnostic count fall (481 → 0). In order:

• let annotations, container literals, statics, pub(read) — let x: multi Text = [], map<K, V>, ?T; []/[a, b]/{} as expressions; static const/static fn with Cls.fn(...) calls; a ; ends a statement so one-line guard bodies parse.
• try/catch over the trap system (plan 8 Task 5) — VM catch frames (TRY/ENDTRY), unwind-to-handler with the try region's own values released, err_fill for the {code, line, method, msg} record, expression and block catch arms. Uncaught traps unchanged.
• nil + 23 text/container builtins — len, byte_at, print_err, starts_with/ends_with, index_of/last_index_of, substr, trim, to_lower, char_of, parse_int, split/split_ws, join, slice, pop/shift, sort, reverse, remove, key_at/val_at, multi_set.
• for k, v in m over a map, and m[i] = v for a multi.
• the systems stdlib's OS half (runtime/src/sysio.c) — fs, time, env, net, proc behind the reserved module names, with predeclared Stat, TimeParts and Proc records and the new WO_T_IO trap.
• json (runtime/src/json.c) + .wob v2 — per-field names, referenced classes and element kinds in the class table, so encode/decode are one metadata-driven implementation; json.decode(t) as T is the language's only cast, yielding ?T.
• program mode — fn main(args: multi Text) -> Int, argv delivered by the runtime, return value as the exit code.
• two safety fixes found by running it: + on Text was lowering to ADD on two heap pointers (now WO-E201 pointing at ..; seven sites in the sample were corrected), and x == nil was lowering to EQS, which dereferences the zero word (now EQ).

Gates at the end of that session: corpus 71/0, woc runtest 565/0, every wovm unit gate green in both dispatch flavors.

---

Done

Language track — compiler + VM (OOP track)

Status	Item	Doc	What actually landed
✅	iteration 37 — wo-html components	37	Two halves. Grammar (2026-08-24): the backtick raw text literal — content verbatim, common margin removed at lex time, ${ } raw and {{ }} auto-escaping to a call on the esc in scope; WO-E004/WO-E005 added; lexer + one parser desugar only, nothing downstream. Library (2026-08-25): Component/render_all/Layout in wo-html, ok_html moved into framework/http beside ok_text/ok_json, site migrated onto Layout + a reused ChapterNav, shop onto AppShell + multi Component children with queries in the controllers; the site was then restructured onto the program template's MVC layout (model / layout / view modules / one controller per feature / bootstrap-only main). multi Component needs no wrapper record — the framework's Mw/Aw shape is not a language requirement. Gates: just site 11/0, just web-app 46/0, woc-test 556/0, oop-e2e 116/0
✅	Principles	../00-principles.md	13 principles, each with a why and a link to the doc that enforces it
✅	wovm VM core	plan 1	.wob v1 loader with full static validation, register interpreter (computed-goto + ISO-C fallback), arena with size-class free lists, borrow word, RC + budgeted Bacon–Rajan cycle collector, drop-map trap unwinding, containers, builtins, ICALL, CLI. 13 suites × 2 dispatch flavors + CLI smoke, ASan/UBSan clean
✅	.wob format contract	oop-vm/00-wob-format.md	Normative; twinned with runtime/src/wob.h
✅	woc compiler front	plan 2	Tasks 1–8: dune scaffold, diag (WO-E codes, two-site related errors, ordered dedup), newline-significant lexer at rt parity, declaration + statement/expression parser with skip-on-block and multi-error recovery, typechecker (field kinds, ?T plumbing, W201, E225, E214), MVS ownership pass with the four emitter tables, driver with directory discovery + cross-file programs. 14 + 264 checks
✅	Error catalog	oop-vm/01-error-catalog.md	14 emitted codes + 10 reserved, each with the reason it is not yet emitted
✅	log-watcher .wo sample	../examples/log-watcher/	Eight-file port authored docs-first with its .hx mapping table; compiles for real in iteration 7
✅	Scalar cleanup	discarded.md	Money/SKU/Float and the abstract allowlist removed; abstract flipped adopt → reject
✅	woc emitter, corpus, single binary	plan 3	Tasks 1–6 + 8 (Task 7, a parity harness against the Rust runtime, deferred by explicit user decision — the two stacks diverge by design). Bytecode emitter (emit.ml) + disassembler (disasm.ml, --dump-bc); three-kind conformance harness (scripts/oop-e2e.sh, just oop-e2e) over tests/corpus/{run,compile-fail,trap,gc}; pricing-demo + ownership/trap corpora (19 fixtures); @gc cycle collector's post-exit pump (WO_GC_BUDGET/WO_GC_TRACE) + 2 gc fixtures (gc/held-cycle retired — see criterion-3 closure below); woc build single-binary output + relocation/corrupt-trailer smoke; WO-E405 closing criterion 3's ASan leak (entry must return Int); just oop-accept wiring all five spec criteria + both unit gates into one command. 14 + 399 compiler checks; oop-e2e 25/25 against the release wovm. Milestone-1 acceptance gate is fully green — all five criteria met (see the dated acceptance note in docs/superpowers/specs/2026-08-01-oop-compiler-vm-design.md)

Known gaps carried out of iteration 3 — recorded, not silently owed:

• ?T is plumbed but unenforced. Lexer/token/AST/parser/dump all handle ?T; the semantics do not exist (WO-E211/E212/E213 declared, never emitted — a probe returning ?Int as Int exits 0). Owned by iteration 5, plan 8 Task 6, which is that iteration's first task because it blocks the log-watcher port. See compiler/nullable-types-implementation.md.
• Structural interface satisfaction is not checked (WO-E205 dead), along with type mismatch, bad arity, and unknown-fn (E201/E203/E204) — all named in plan 2 Task 6's own must-fail list. Gaps in shipped work, catalogued as reserved.
• Six further narrowings (W201 heuristic, E225 reach, dead code after return, unresolved-callee drops, RC table ordering, residual b-side role) are listed in the plan-2 SDD ledger and in the affected files' own comments.

Known gaps carried out of iteration 4 — recorded, not silently owed:

• WO-E205 (unsatisfied interface) reachable but unenforced — closed 2026-08-18 (branch type-enforcement): structural satisfaction is checked at call arguments, annotated lets, and returns; the pinned fixture moved to compile-fail/unsatisfied-interface with fixture.code WO-E205 in the same change, as its comment demanded. The hybrid boundary is restored.
• set(m, k, v)'s @gc retention gap on map keys/values is open — the twin of the push bug Task 5 fixed for multi. set has no equivalent special case in owner.ml's analyze_call, so a @gc key or value handed to set is under-counted and the collector can free it while the map still points at it. Nothing in the corpus exercises this yet. See oop-vm/08-builtin-surface.md.

Known gaps carried out of the 2026-08-14 compile-and-run milestone — recorded, not silently owed:

• Optionals are lenient. ?T has its representation (the zero word) and its comparisons, but WO-E211–E213 are still dead: a ?T may be used where T is required, and nothing narrows inside an if x != nil branch. The workload leans on that leniency today.
• pub(read) is parsed, not enforced. The marker rides on the field (Ast.field.pub_read); no check refuses a write from outside the declaring class yet.
• using extensions and #if build flags are absent, and the reject rows (extends/cast/Dynamic/…) still have no doctrine-citing diagnostics — plan 8 Tasks 7–8's remainder.
• A borrowed non-constant Text pushed into a container is a double-free hazard — closed 2026-08-14 by copy-on-push: push/set/m[i] = v copy a TEXT element, key or value into the container, and the compiler drops a freshly built Text right after the call (a value read out of a place keeps its owner). The failure it fixed was real: a tools/call of tail_log used to answer {"isError":true,"text":"tool failed: not a text value"}; all four MCP tools now return isError:false with correct payloads. OWNED/GCREF elements still move, and set's @gc retention gap is still open (see oop-vm/08-builtin-surface.md).
• A blocking accept/read swallows SIGTERM. env.stopping() installs a handler that only sets a flag, and net.accept/net.read retry on EINTR, so a server parked in accept never observes it: a plain TERM does not stop the process (timeout -k / kill -9 does). Graceful shutdown needs an interruptible wait — the shard-actor runtime's event loop (iteration 8) is where that belongs, not a patch to the blocking calls.
• A temporary record whose field is iterated is never dropped — for e in parse_dir(dir).entries keeps the entries alive (good) but leaks the ParseResult shell (its drop is recorded for no register). Found in the same disassembly; a leak, not a corruption.
• json's two documented limits — closed 2026-08-18 (branch json-fidelity): a Bool field encodes true/false (WOB_FIELD_BOOL / WOB_FIELD_NIL_BOOL in the field metadata), and a fraction/exponent is malformed for an Int field — the checked decode yields nil instead of truncating (floats stay representable via a raw json.Value field).
• net fd lifetime is the program's problem. net.close exists; the sample's MCP server never calls it, so a long-running mcp session leaks descriptors. That is the sample's bug to fix, not the runtime's.
• The workload has never run under ASan, and iteration 4's gc/held-cycle leak (above) is still open. The corpus itself stays ASan-clean.
• json.encode Bool/nil-scalar asymmetry — closed 2026-08-18 with the same change: Bool encodes true/false, ?Bool nil encodes null.
• No corpus fixtures cover the new surface. By explicit direction (2026-08-14) the acceptance for this work is the log-watcher program itself, not fixture pairs; tests/corpus/ still gates every pre-existing behavior (71 checks, 0 failures).
• E201/E203 and seven other WO-E2xx codes remain declared but unemitted — see oop-vm/01-error-catalog.md.
• CLOSED — milestone-1's ASan gate (just oop-accept) failing on gc/held-cycle. Root cause (Task 8's finding, restated): main.c's entry-method return value (uint64_t ret, src/main.c:158) is stored but never released, so gc/held-cycle's "permanent external hold" was actually a permanent refcount inflation — LeakSanitizer's "definite leak" (1184 bytes / 3 allocations) was correctly reporting exactly that, not a false positive. Fixing it by releasing ret was rejected: the .wob method table carries no return-type/kind metadata, so main.c has no way to know ret is a pointer rather than a scalar, and adding that metadata is a format change out of scope here. Fixed instead at the source: the systems-track spec already requires the entry to return Int (its return value is the process exit code), so a class-returning main was never legal — WO-E405 (compiler/src/emit.ml, 01-error-catalog.md) now rejects it at compile time, and gc/held-cycle is retired because its premise (an externally-held cycle survives a post-exit pump) is no longer expressible — see oop-vm/02-corpus.md's "Retired" note for why, and for where the scenario it meant to cover is actually proven (runtime/test/test_cycle.c, plus a proper in-flight fixture scheduled for story iteration 7b). Spec success criterion 3 is now MET; just oop-accept passes all five criteria.

The C proving-ground work (exploration/c-runtime/, phases A–F: 859k reads/s, 618k durable commits/s) fed the current C runtime and remains as an exploration study.

---

Pending

Measured optimization candidates live in docs/plan/perf-targets.md — a register like discarded/learnings: a target enters with a number, leaves by landing (baseline delta) or by rejection into discarded.md.

Implementation order (re-sequenced 2026-08-21 — concurrency chain)

Everything still pending IS the runtime-concurrency chain. Basis: the 2026-08-20 code-review pass (measure before optimizing, close correctness holes before adding surface), amended 2026-08-21 by developer decision: stage 3 before 22 — correctness first, then one benchmark campaign covers single- and multi-shard. The authoritative table with per-row reasoning is 00-story.md.

Dependency rules that force the shape: 23 after stage 3 + 22 (the ring is the arc's, the baseline is 22's); 24 after 31 (chat is dishonest without lifecycle); h2c stays parked behind the chain; the held tail keeps its own precedence notes for resumption.

1. ✅ 8+11 stage 3 — landed 2026-08-21 (just db-actor 8/0; arc complete, stories in done/). Was: transparent DB actor. A correctness fix, not an optimization: worker VMs are zero-initialized, so a DB statement off the primary traps WO_T_DB — a multi-shard program touching the database is broken today. Plan of record: 2026-08-20-shard-fiber-arc.md (stages 1+2 landed 2026-08-20, branch concurrency-arc).
2. ✅ 22 — LANDED 2026-08-21 (just db-bench, baseline committed, gate bites; headline: durable 4.5k vs ram 297k inserts/s, reads O(table), mixread 21 ops/s multi-shard, msgrate 2.45M cross-shard). Was: the measurement backbone: restart-persistence proof + baseline benchmark (durable + RAM-only), single- AND multi-shard in one campaign, plus the stage-2 mutex-inbox number (rings only if the mutex costs). It has never run — no bench/baseline.json, no just db-bench; the arc's stages 1+2 delta is recorded retroactively.
3. 31 — actor lifecycle (story, written 2026-08-21): request/response (send is one-way and callers sleep to await), bounded mailboxes (the FIFO only grows), actor death/supervision, timers beyond time.sleep.
4. 24 — chat, the arc's acceptance; honest only after 31 (19 landed 2026-08-20 — Bytes carries the frames).
5. 23 — io_uring group-commit; the WAL's WRITE+FSYNC chains ride the arc's per-shard ring (T4); after 22's baseline — the payoff, measured.
6. 32 — WAL checkpoint (story, written 2026-08-21): the WAL is append-only forever — snapshot + truncate reclaims disk and bounds replay; after 23 (composes with group-commit), policy set by 22's aged-store numbers.

30 — observability, CI, fuzz: named 2026-08-20, still row-only (no story file); slots in when scheduled — nothing in the chain depends on it.

▸ databasev2 — the database beyond RAM

New 2026-08-26. The problem: rows were resident unconditionally and nothing declares a budget. Rows live in malloc'd slabs whose addresses are stable forever; there is no eviction, spill or paging anywhere in database/src/; the WAL never checkpoints so boot replays all history; and durability is one process-global WO_DATA, so no table can say it matters more than another. An allocation failure is a clean catchable WO_T_OOM only in the VM arena — table storage has no ceiling and is SIGKILLed instead (measured, databasev2 1). Where swap exists the ceiling may never announce itself at all: an append-mostly 900k-row run finished at uncapped speed inside a 64 MiB cap (148 s vs 150 s), serving from disk with no error signal.

The lever is per-table storage modes, which is why this track has a grammar iteration. Six pending iterations moved here from the language track (their old ids in the rows below); four are new. Done database work — 9, 9b, 22 — stays in the language arc as v1 history.

#	Iteration	State
1	RAM ceiling: measure the breaking point	✅ MEASURED 2026-08-27 — readiness: ready, status: done; forks settled, harness landed (148 checks). Footprint 96.5–100 B/row Int vs 320.6–324 B/row text = 3.3× (not the "order of magnitude" three docs claimed), read as median-of-marginals because doublings swing a two-point slope 2×. Both predicted exits were wrong: table storage has no checked ceiling and is SIGKILLed (overcommit lets malloc succeed, kernel kills on page touch), and swap is not latency collapse — 900k rows finished 148 s capped-with-swap vs 150 s uncapped, ~1%, returning 0 while serving from disk. Ack-after-fsync survives an OOM kill: ~40 000 rows recovered as an intact prefix, gated as the ceiling leg. Also measured: random reads over an oversized table collapse 273× (1.85M vs 6 771 reads/s, p99 1 µs vs 487 µs) — so the two access patterns sit ~270× apart under the same pressure, and departure is a step, not a curve. Replay measured too: ≈5.5 µs/record, 1.9× history penalty (10M records ≈ 55 s of boot) — iteration 3's missing "before", now gated. Iteration 2's budget dependency is removed, not satisfied — there is no "swap onset" to derive it from
2	per-table storage: durable and resident	🔄 the language enrichment — the durable half is DONE and usable. Two optional @table keys, durable: true|false and resident: all|keys, both defaulting to today's behaviour (all 28 existing declarations compile unchanged, no golden moved). Landed: the grammar, WO-E224 (a durable ref into a volatile table is refused), .wob v7 carrying both properties in spare flags bits, durable: false actually skipping the WAL (measured: 50 inserts → 1500 bytes durable, 0 volatile) with a mode-mismatch startup refusal, plus offset capture and read-a-row-from-an-offset. Outstanding: 5c/5d (the id→offset map and rewiring wo_row_ptr's 11 call sites, slab scans and @unique/FK across the boundary — not yet written up), the two runtime refusals, and closeout. spec · plan
3	WAL checkpoint (was 32)	✅ LANDED 2026-08-29 — the chain's last link. Compaction rewrites the log as one record per live row and swaps it in with rename, so recovery is completely unchanged and crash safety comes from the filesystem rather than from code. 2.16× space reclaimed (1 962 358 → 907 094 B), boot 114 → 64 ms, stop-the-world pause 2 651 µs against a stated 50 ms budget. Read .dev/reference/postgresql for it: PG never compacts its WAL — its records are page deltas, so it needs heap files, a control file, a redo pointer and a separate process. Ours are full row images, so a compacted log IS a store, which deletes all of that. kill -9 during compaction: 40 rounds/run, 10 clean runs, and mutation-proven — against in-place rewrite instead of rename the battery fails every time. Outstanding: the resident: keys offset map (compaction moves every record; the obligation is recorded at the compactor) and the O(live rows) pause, ~5.5 s at 1 GB, which is what an incremental design must be bought against
4	io_uring group commit (was 23)	✅ part A LANDED 2026-08-28 — group commit, one barrier per drain instead of one per statement (the engine was fsync-per-STATEMENT, not per commit; the story's premise was wrong). Shard 0 holds each reply, commits once when its queue empties, releases all — so a writer is acked after the barrier carrying ITS record. ≈2.9× durable write throughput, ≈2.1× lower p50, two measurement methods agreeing (2.9× controlled, 3.5× s1-vs-sN); mean batch 5.43, peak 57. A durability failure is now fatal (exit 74), not a catchable WO_T_IO — replacing three behaviours that disagreed, two of which admitted leaving RAM ahead of disk. What it did NOT do: durable.sN.mixwrite 480→492 (unchanged — that workload does 20 writes at C=4, mean batch 1.01) and seed unchanged (serial writers have nothing to batch with). This row used to say "close the 66× gap"; that target was mis-stated — the gap is two problems and part A fixes only the concurrent one. ⬜ part B (io_uring) needs re-brainstorming, not starting on the old premise
5	Bounded tables and eviction	⬜ a declared capacity + refuse/evict/back-pressure, and a process-level pressure signal that sheds before the allocator or OS gets involved — turning the invisible failure into a managed one
6	Cold tiering	⚠ largely superseded by 2 — resident: keys took the ceiling-raising role; its user-space-working-set premise was rejected for the kernel page cache. Mostly forks: which shape, whether the index itself fits, whether the language surfaces the fault cost, and whether @unique on a cold table is refused outright. A paged B-tree stays rejected — if tiering needs one, reject tiering
7	Single-file store (was 33)	⬜ WO_DATA=<path>.db; driver-only, independent
8	Query grammar from corpora (was 27)	⬜ whole-query count, exists; independent
9	Cross-program tables (was 20)	⏸ hold — attach to a running program's database over local IPC
10	Keypair attach auth (was 21)	⏸ hold — program identity as a keypair; needs 9
11	Bounded delta chains	✅ LANDED 2026-08-30. A resident: keys row's delta chain is bounded in the UPDATE path, because the checkpoint is blind to per-row chain length — it thresholds on whole-log bytes, so one hot row can grow an unbounded chain inside a log that never trips compaction. The fold now reports hop count (free — the walk already visited every hop), and past WO_DELTA_MAX_HOPS (16) the update writes a full row image instead of a delta, resetting depth to 0. Two things the tests corrected. The flattened image is a WO_WAL_UPDATE, not an INSERT: the row's original INSERT is already in a live log, so a second one for the same id is a duplicate that replay correctly refuses as corruption — INSERT is right only for compaction, which builds a fresh log. And the proportional ceiling was removed as dead code: with the absolute term at 64 MiB, garbage large enough to reach a 256 MiB ceiling has already tripped it, so the branch was unreachable. Borrowing both constants from postgres was the wrong inference — PG needs two because it thresholds on tuples with its pair at opposite ends (base 50, max 1e8); this thresholds on bytes, where one constant does both jobs. Found by trying to write a test for the ceiling and finding no input could reach it. Four tests: depth stays bounded across 2K+2 updates, a flattened chain replays, a delta on an indexed column composes with flattening (checked at every step across the bound and after restart — found no product defect), and the policy's absolute term with its boundary. test_wal 5700 pass / 0 fail; wovm-test and woc-test green. One criterion is weaker than written: the replay check asserts an expected value, not a resident: all oracle table. spec
12	Schema migrations	✅ LANDED 2026-08-31. A @table class is the schema, the log is the database, and boot now compares them — before this, an added or deleted field turned a healthy WO_DATA into "corruption" and reordering declarations silently decoded rows into the wrong class. Landed: WO_WAL_SCHEMA head record (written LAZILY ahead of the first real record — an eager head broke durable: false's documented zero-bytes contract by 75 bytes and the gate caught it), a name-keyed diff whose refusals are per-class POISONS that bite only when a record of the class is met, and a record-level TRANSCODE: cids remap by name including inside stored owned values, deleted values freed, added fields zero-filled, delta back-pointers rewritten through an offset map with deltas on deleted fields SPLICED out; temp+fsync+rename, compaction's crash discipline. Two bugs the tests forced out: a poisoned class skipped plan identity so the retype refusal fell through to generic "corruption" (the message this iteration exists to replace), and early goto corrupt freed uninitialized memory. End-to-end: migrating \Note`: +flagthenflag=0; retype refuses naming val, exit 2, old binary still boots the refused log. 21 new tests, test_wal **5966/0**; wovm/woc/site/residency gates green. v2 holds rename (@renamed_from`), retypes, and data/seed migrations. spec

---

▸ porch — the web framework track

New 2026-08-26, from the Fiber v3.5.0 parity study. Supersedes language iteration 39, now a pointer. All eight are ⬜ refine — none has an approved spec yet. Ordered by dependency; the first slice is deliberately the cheapest so the store pattern and gate shape are proven before the runtime and Resp are touched.

#	Iteration	State
1	Store-backed middleware	⬜ startable today — rate limiter + idempotency over a @table; needs no new primitive, only time.ticks. Durable counters are the differentiator over Fiber's in-memory default, so the gate includes a restart
2	Randomness and cookies	⬜ the foundation. Phase A is language-track work: a CSPRNG builtin (id 96+; 89/90 are iteration 31's reserved holes). Then repeated response headers — Resp.headers is a map<Text,Text> and structurally cannot emit two Set-Cookie lines — then Cookie: parsing and signed cookies
3	Sessions	⬜ after 2. Server-side rows keyed by a random id, idle and absolute timeout, id rotation on login, revoke-all-for-principal, durable across restart
4	CSRF	⬜ after 2 + 3. Session-bound tokens, trusted origins as the second layer, opt-in single use, and refusal classes that are distinguishable in logs
5	Routing + response ergonomics	⬜ independent, any time — patch/options/head/all, named routes + URL building, per-route body limit (today BODY_MAX is one compile-time number), request ids, Location/Vary/Attachment, and q-value ranking (retires a standing 🔶)
6	Streaming core	⬜ the riskiest and highest-leverage slice: incremental writes + chunked framing + an explicit commit point. serialize() always emits Content-Length today. Chunked REQUEST bodies are deliberately refused (request smuggling) and that refusal must survive
7	SSE + compression	⬜ after 6. SSE fits the actor/fiber model unusually well; compression carries a real fork — pure-.wo DEFLATE (now expressible after iteration 36's bit operators) vs a C builtin. CRC32 finally gets its consumer
8	Static files + lifecycle	⬜ static half after 6. Byte ranges, Last-Modified/Cache-Control, index resolution, listing off-by-default, shutdown hooks (the ledger's "no user teardown hooks yet"), plus healthcheck/favicon/redirect/rewrite/skip

---

⏸ Held (2026-08-21, developer decision): 18, 20, 21, 25, 26, 27, 28, 29 — every story carrying status: hold in its frontmatter (25's story file removed; its plan doc remains). Half-done branches (ipc-attach, keypair-auth) keep their manifests.

✅ 17 — landed 2026-08-20 (unparked and executed): kind = "library", check mode, and the internal/ dep boundary (WO-E108). Driver-only. ✅ 19 — landed 2026-08-20: Float + Bytes, .wob v5.

Language track — sequenced, on the critical path

#	Item	Plan
5	Haxe-parity language surface — ?T forced handling first, then switch expressions, records, enum payloads, try/catch, statics, using, modules, is, pub(read), #if	plan 8
6	Program mode + systems stdlib — fn main, exit codes, fs/proc/net/time/json	plan 9
7	log-watcher proof — the sample compiles and detects a silent death live	plan 10
8	Shard-actor runtime	arc plan (plan 4 ✖ discarded 2026-08-21 — epoll-based)
9	Database engine binding	plan 5
9b	@table + relations + language-integrated query — comprehension queries, ref/backlink navigation, GroupBy aggregates; acceptance: new docs/examples/employee sample	spec · plan
20	Cross-program tables — attach to a running program's database (IPC string in wo.toml, manifest-granted rights, owner stays the single writer)	no spec yet — four open forks recorded in the iteration; brainstorm before planning
21	Keypair attach auth — mutual challenge–response, grants name public keys, uid superseded	no spec yet — four forks recorded; plan folds into 20's
22	Durability + throughput + scale — restart-persistence, read/write benchmark, ~1M rows; the gate every later optimization re-runs	no spec yet — four forks recorded; the measurement backbone
23	io_uring group-commit write path — batched durability overlapped on shard threads, fsync fallback	no spec yet — brainstorm after iterations 8 + 22
27	Query grammar from real embedded-DB corpora — whole-query count + correlated exists, driven by the skillhost SQL catalogue; add only what a corpus uses	no spec yet — three forks; may collapse to "confirm len(query) + add exists"
14	skillhost host workload — port skillhost (MCP host + confined script runner) to writeonce; drives the missing host capabilities into the open (bounded subprocess, stdin/stdout transport, fs metadata, FFI-vs-out-of-process)	no spec yet — gaps recorded in the iteration; each gap brainstormed on demand, bounded-subprocess first
17	library projects + dependency privacy — wo.toml kind = "library" (checkable without entry, dual lib+bin) + Go-style internal/ at the [deps] boundary; framework reorg demonstrates both	✅ landed 2026-08-20 — spec · plan
10	HTTP service layer	plan 6
11	Fibers	vision §3, blue-green exploration
12	Blue-green deploy	spec — plan authored after iterations 9 + 25

Language track — parked until after iteration 26

Recorded 2026-08-08 by scope directive; nothing here lands before the log-watcher proof.

• WO-W201 @gc-suggestion refinement beyond the self-reference heuristic
• WO-E225 broadened to ref/multi/map element types and fn signatures
• ADT container roster adoption (Stack, Queue, Set, Tree, Graph, …) — see the roster in compiler/nullable-types-implementation.md
• Web framework as a .wo library; UI (##ui SSR + live patches); script-based destructive migrations; MCP/agent wrapper over the management plane
• throw (explicit raise) — cut 2026-08-10, 0 uses in the driving workload (log-watcher); catch frames ship without it
• time.mono — cut 2026-08-10, 0 uses in the driving workload; returns when a workload needs monotonic math
• is — cut 2026-08-10, 0 uses in the driving workload; emptied plan 8's old Task 7, which is deleted rather than deferred

Frontend — removed as stale (2026-08-17)

The ##ui / .htmlx LiveView frontend track — 13d pricing UI, the 14-MVC-UI implementation plan, the 7-of-7 ui-htmlx-live plan, and the 9-doc plan/exploration/ui/ design set — was removed. It was built entirely on the non-advancing Rust runtime (.dev/reference/crates/wo-htmlx, cargo run, WebSocket live-patches) and contradicts the current woc/wovm direction. Recorded in discarded.md.

---

Discarded

Settled rejections with their reasons live in discarded.md — inheritance, abstract newtypes, Money/SKU/Float, Dynamic/cast/ macro/extern, AOT-to-C, Menhir, shared mutable engine state, external deployer daemon, destructive migrations in v1, and more. Argue against the recorded reason rather than re-opening an entry as new.

Learnings

What attempts taught, shipped or not, in learnings.md — plumbed-is-not-enforced, vacuously-passing goldens, exit-0-with-wrong-output, the malloc-path ASan trick, deferred checks that never reach the runtime, validate-once-at-the-boundary, and reference-implement-in-C-first.