fix: net.Conn is a scalar, \r escape, map[k] is optional, interp node ids

Found by driving the compiled log-watcher's third path — the MCP server. It now
answers real JSON-RPC over HTTP: initialize returns protocolVersion/serverInfo,
tools/list returns the full tool list (1049 bytes of generated JSON), and an
unauthorized request gets 401 {"error":"unauthorized"}. corpus 71/0, woc 565/0,
wovm gates green.

- lexer: `\r` and `\0` escapes. Without `\r` a program cannot write CRLF at
  all — the server's `index_of(buf, "\r\n\r\n")` was searching for a literal
  backslash-r, so it never found a header terminator and hung on every request
- parser: an interpolated sub-expression now mints node ids from the OUTER id
  space. A fresh sub-parser started at 1, so `${...}` nodes collided with the
  file's own nodes — and every side table (drops, moves, rc, masks, f_decl) is
  keyed by node id. Surfaced as WO-E404 "ownership table names `headers`,
  which has no register"; silent misattribution otherwise
- types.ml: `net.Conn` is a reserved SCALAR type (a file descriptor). It was
  falling through as "some user class", i.e. WO_K_OWNED, so the frame would
  DROP an integer at scope end
- types.ml: confident_typ knows `..` yields Text. Interpolation desugars to a
  Concat chain, so without it every interpolated value looked underivable —
  which is why the `+`-on-Text check missed two live sites in the workload
- `m[k]` on a map is now the OPTIONAL read (nil for a missing key), while
  `get(m, k)` stays the asserting one that traps KEY. That is what makes
  `let v = m[k]; if v != nil` — the workload's header lookup — work.
  trap/missing-map-key now pins `get(...)`, and the surface doc records the
  split
- json.encode of a `json.Value` emits it verbatim (kind 255): an echoed id was
  coming back as "1" instead of 1
- disasm: TRY/ENDTRY render instead of ?OP32/?OP33
- status board: the push-of-a-borrowed-Text gap is now recorded with the
  concrete failure it produces (tools/call tail_log), plus the leaked
  temporary-record shell found in the same disassembly

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
shoney.arickathil 2026-08-14 17:47:45 +02:00
parent 97d1d3024e
commit ffb8b3b952
13 changed files with 119 additions and 19 deletions

View file

@ -147,6 +147,10 @@ let ins_str (i : int) (pc : int) : string =
else Printf.sprintf "BUILTIN r%d, r%d, %s" a b (builtin_name c)
| 30 -> "DB_STUB"
| 31 -> Printf.sprintf "TRAP %d" bx
(* haxe-parity Task 5: try/catch. The handler target is rendered the way
jumps are — absolute, so a disassembly can be read against the pc column. *)
| 32 -> Printf.sprintf "TRY r%d, handler -> %04d" a target
| 33 -> "ENDTRY"
| op -> Printf.sprintf "?OP%d" op
(* ---- the dump ---- *)

View file

@ -255,6 +255,7 @@ let b_map_remove = 36
let b_map_key_at = 37
let b_map_val_at = 38
let b_multi_set = 39
let b_map_get_opt = 59
(* json (runtime/src/json.c): encode takes the value's static kind as its
second argument, decode the class id to build as its second. *)
@ -1518,7 +1519,12 @@ let rec emit_expr (p : pctx) (f : fstate) (v : views) ~(dst : int) ?expected (e
| Index (base, idx) -> (
let bid =
match ty_of_expr p f base with
| Some bt -> ( match unwrap bt with Multi _ -> Some b_multi_get | Map _ -> Some b_map_get | _ -> None)
(* `m[k]` on a map is the OPTIONAL read — a missing key is nil, which is
what makes `let v = m[k]; if v != nil { ... }` the ordinary lookup
idiom. `get(m, k)` keeps asserting (WO_B_MAP_GET traps KEY). A `multi`
index still traps out of range: a bad index is a fault, not an
absence. *)
| Some bt -> ( match unwrap bt with Multi _ -> Some b_multi_get | Map _ -> Some b_map_get_opt | _ -> None)
| None -> None
in
match bid with
@ -2446,7 +2452,13 @@ and emit_call (p : pctx) (f : fstate) (v : views) ~(dst : int) ?expected (e : As
emit_expr p f v ~dst:base a;
f.f_temp <- save;
let kind =
match ty_of_expr p f a with Some t -> field_kind p t | None -> 3 (* Text *)
match ty_of_expr p f a with
(* a `json.Value` is raw JSON already: kind 255 tells the
builtin to emit it verbatim instead of quoting it *)
| Some t when (match unwrap t with Scalar n -> n = Types.json_value_type | _ -> false)
-> 255
| Some t -> field_kind p t
| None -> 3 (* Text *)
in
put f (ins_abx op_loadk (base + 1) (check_bx p f e.pos "constant" (const_int p kind)));
sync_mask p f v e.id;

View file

@ -10,7 +10,8 @@
newline;
- single- or double-quoted strings support the same backslash
escapes as rt: n, t, backslash, or either quote character, each
backslash-prefixed; anything else verbatim;
backslash-prefixed; anything else verbatim. `\r` and `\0` were added
2026-08-14 (a program writing HTTP needs CRLF, and rt never had to);
- integer literals are plain runs of ASCII digits;
- identifiers may contain internal dashes, exactly like rt's
read_ident_chars (crates/rt/src/lexer.rs) — so `foo-bar` lexes as
@ -291,6 +292,12 @@ let tokenize (collector : Diag.Collector.t) ~(file : string) (src : string) :
match advance lx with
| Some 'n' -> Buffer.add_char buf '\n'
| Some 't' -> Buffer.add_char buf '\t'
(* `\r` — added 2026-08-14: without it a program cannot write CRLF
at all, and the driving workload's HTTP server needs it (its
`index_of(buf, "\r\n\r\n")` was searching for a literal
backslash-r, so it never found a header terminator). *)
| Some 'r' -> Buffer.add_char buf '\r'
| Some '0' -> Buffer.add_char buf '\000'
| Some '\\' -> Buffer.add_char buf '\\'
| Some '"' -> Buffer.add_char buf '"'
| Some '\'' -> Buffer.add_char buf '\''

View file

@ -1121,6 +1121,15 @@ and desugar_interp (st : state) (pos : Ast.pos) (segs : Token.str_part list) : A
let sub_collector = Diag.Collector.create () in
let sub_toks = Lexer.tokenize sub_collector ~file:st.file raw in
let sub_st = make sub_collector ~file:st.file sub_toks in
(* The sub-parse must mint ids from the OUTER id space and hand back what
it used. A fresh state starts at 1, so every interpolation used to
produce nodes whose ids collided with unrelated nodes of the same file —
and every side table downstream (drops, moves, rc sites, masks,
f_decl) is keyed by node id, so a collision silently attributes one
construct's ownership table to another. That surfaced as a WO-E404
("ownership table names `headers`, which has no register") on a method
whose interpolation happened to collide with a later local's node. *)
sub_st.next_id <- st.next_id;
let parsed =
try
let e = parse_expr sub_st in
@ -1128,6 +1137,7 @@ and desugar_interp (st : state) (pos : Ast.pos) (segs : Token.str_part list) : A
else None
with Parse_error -> None
in
st.next_id <- sub_st.next_id;
match parsed with
| Some e -> e
| None -> fail st pos syntax_code "malformed \"${...}\" interpolation expression"

View file

@ -179,6 +179,16 @@ let is_stdlib_module (name : string) : bool = List.mem name stdlib_modules
dynamic value tree. *)
let json_value_type = "json.Value"
(* The reserved stdlib TYPE names and their representations. `net.Conn` is a
file descriptor — a SCALAR — and getting this wrong is not cosmetic: an
unknown qualified name would fall through to "some user class", i.e.
WO_K_OWNED, and the frame would DROP an integer at scope end. Anything
qualified with a stdlib module and not listed here is an error at the use
site rather than a guess (types.ml's check_use_edges). *)
let stdlib_scalar_types = [ "net.Conn" ]
let is_stdlib_scalar_type (name : string) : bool = List.mem name stdlib_scalar_types
(* haxe-parity Task 5: the record `catch (e)` binds — the VM's structured
trap error, one shape forever (spec §6). Predeclared rather than
written: no source declares it, every program that catches gets it, and
@ -357,6 +367,7 @@ let wob_kind_of_typ (syms : symbols) (t : typ) : wob_kind =
came from, which json.encode emits back verbatim. Kinding it TEXT
is what makes it drop correctly and pass through concatenation. *)
if name = "Text" || name = json_value_type then WO_K_TEXT
else if is_stdlib_scalar_type name then WO_K_SCALAR
else if is_builtin_scalar name then WO_K_SCALAR
else if StringMap.mem name syms.unions then
(* haxe-parity Task 4: an all-bare union value is a plain
@ -1130,6 +1141,11 @@ let typecheck_program ~file ~(module_of : string -> string)
`a == 1 and b == 2` without a false "underivable" silence on
the left-hand comparison. *)
Some (TScalar "Bool")
(* `a .. b` is CONCAT, which always produces a fresh Text — and string
interpolation desugars to exactly such a chain (parser.ml), so without
this arm every interpolated value looked underivable and every check
built on confident types silently skipped it. *)
| Binary (Concat, _, _) -> Some (TScalar "Text")
| Interp _ ->
(* An interpolation always *produces* Text by construction
(emit.ml decides, per-segment, whether the embedded value

View file

@ -181,10 +181,21 @@ recorded, not silently owed:
(`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** — `push(m, v)`'s open gap, now shared by list literals (`[a, b]`)
and documented in `owner.ml`'s own comment. The workload's literals are
string constants or borrowed params handed straight to a stdlib call, so
nothing reachable today hits it.
hazard** — `push(m, v)`'s open gap, now shared by list literals (`[a, b]`).
**Reached for real on 2026-08-14**: the MCP server answers `initialize`,
`tools/list` and an unauthorized request correctly, but a `tools/call` of
`tail_log` returns `{"isError":true,"text":"tool failed: not a text value"}`
— `Mcp.allowed_paths` pushes `e.log_path` (a Text the entry record owns)
into a fresh `multi Text`, so two owners free one string and a later read
finds recycled memory. The fix is a coordinated pair, not a one-liner:
`multi_push`/`map_set` must COPY a TEXT element (as `slice` already does),
and `owner.ml` must then stop treating a pushed Text as escaping so the
caller's own fresh temporaries are still dropped. Until then, treat
`push`-of-a-borrowed-Text as unsound.
- **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**: a `Bool` field encodes as `0`/`1` (the
class-table kind byte does not distinguish it from an integer), and a JSON
number with a fraction or exponent decodes by truncation.
@ -193,6 +204,9 @@ recorded, not silently owed:
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` of a `Bool` and of a nil scalar are asymmetric**: a nullable
scalar encodes as `null` (the field metadata says so), a plain `Bool` still
encodes as `0`/`1`.
- **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

View file

@ -126,7 +126,7 @@ class Mcp {
let resp = self.handle(req);
let head = "HTTP/1.1 ${resp.status} ${status_text(resp.status)}\r\n";
head = head .. "Content-Type: application/json\r\nConnection: close\r\n";
head = head + "Content-Length: ${len(resp.body)}\r\n\r\n";
head = head .. "Content-Length: ${len(resp.body)}\r\n\r\n";
try net.write(c, head .. resp.body) catch (e) {}
}
}

View file

@ -52,9 +52,18 @@ maps to one `BUILTIN` id of the format doc.
given, so one source name covers the `multi` and `map` ids the runtime
keeps apart.
**Sugar.** `c[i]` is exactly `get(c, i)` and `m[k] = v` is exactly
`set(m, k, v)` for a `map` and `multi_set(m, i, v)` for a `multi` (the
element it replaces is the container's, so the VM drops it).
**Sugar.** `m[k] = v` is exactly `set(m, k, v)` for a `map` and
`multi_set(m, i, v)` for a `multi` (the element it replaces is the container's,
so the VM drops it).
Reads differ by container, deliberately (amended 2026-08-14):
- `c[i]` on a **`multi`** is `get(c, i)` — an out-of-range index traps
`BOUNDS`, because a bad index is a fault, not an absence.
- `m[k]` on a **`map`** is the **optional** read (`map_get_opt`): a missing key
yields nil, which is what makes `let v = m[k]; if v != nil { … }` the
ordinary lookup idiom the driving workload uses for HTTP headers.
`get(m, k)` remains the **asserting** read and still traps `KEY`.
**Shadowing.** A user-declared free `fn` of the same name always wins. A
declared name is never silently replaced by a builtin.

View file

@ -63,9 +63,11 @@ static int elem_cmp(uint8_t kind, uint64_t a, uint64_t b) {
int wo_builtin(wo_vm *vm, uint64_t *R, uint32_t ins, const char **msg) {
wo_rt *rt = &vm->rt;
uint8_t A = wo_ins_a(ins), B = wo_ins_b(ins), C = wo_ins_c(ins);
/* the OS half and json live in their own translation units */
if (C >= WO_B_JSON_ENCODE) return wo_builtin_json(vm, R, ins, msg);
if (C >= WO_B_SYS_FIRST) return wo_builtin_sys(vm, R, ins, msg);
/* the OS half and json live in their own translation units; ids outside
both ranges (WO_B_MAP_GET_OPT and anything added after it) stay here */
if (C == WO_B_JSON_ENCODE || C == WO_B_JSON_DECODE)
return wo_builtin_json(vm, R, ins, msg);
if (C >= WO_B_SYS_FIRST && C <= WO_B_PROC_RUN) return wo_builtin_sys(vm, R, ins, msg);
switch (C) {
case WO_B_NOW: { /* wall-clock milliseconds */
struct timespec ts;
@ -185,6 +187,12 @@ int wo_builtin(wo_vm *vm, uint64_t *R, uint32_t ins, const char **msg) {
}
return 0;
}
case WO_B_MAP_GET_OPT: { /* `m[k]`: a missing key is nil, not a trap */
wo_map *m = native_check(R[B], WO_CLS_MAP, msg);
if (!m) return WO_T_BOUNDS;
if (wo_map_get(m, R[B + 1], &R[A]) != 0) R[A] = 0;
return 0;
}
case WO_B_MAP_HAS: {
wo_map *m = native_check(R[B], WO_CLS_MAP, msg);
if (!m) return WO_T_BOUNDS;

View file

@ -548,7 +548,11 @@ int wo_builtin_json(wo_vm *vm, uint64_t *R, uint32_t ins, const char **msg) {
case WO_B_JSON_ENCODE: {
jbuf b = {NULL, 0, 0, 0};
uint8_t kind = (uint8_t)R[B + 1];
enc_value(&b, vm->mod, R[B], kind, WOB_NONE);
/* a `json.Value` argument is raw JSON already (an id echoed back into
a response, say) — quoting it as a Text would change `1` into `"1"` */
uint32_t fclass = kind == WO_JSON_KIND_RAW ? WOB_FIELD_JSON_RAW : WOB_NONE;
if (kind == WO_JSON_KIND_RAW) kind = WO_K_TEXT;
enc_value(&b, vm->mod, R[B], kind, fclass);
if (b.oom) {
free(b.p);
*msg = "out of memory";

View file

@ -61,7 +61,7 @@ static const uint8_t b_arity[WO_B_MAX + 1] = {
[WO_B_NET_CLOSE] = 1, [WO_B_PROC_RUN] = 3,
/* json (json.c): encode takes the value's static kind, decode the class
id to build */
[WO_B_JSON_ENCODE] = 2, [WO_B_JSON_DECODE] = 2,
[WO_B_JSON_ENCODE] = 2, [WO_B_JSON_DECODE] = 2, [WO_B_MAP_GET_OPT] = 2,
};
static int vtab_cmp(const void *a, const void *b) {

View file

@ -273,13 +273,22 @@ enum {
* top level comes from object headers and the class table. decode takes
* the class id to build, and yields nil (0) on malformed input — never a
* trap, which is what makes `json.decode(t) as T` a checked decode. ---- */
WO_B_JSON_ENCODE = 57, /* (value, kind) -> Text */
WO_B_JSON_ENCODE = 57, /* (value, kind) -> Text; kind 255 = the value is a
* `json.Value`, i.e. raw JSON to emit verbatim */
WO_B_JSON_DECODE = 58, /* (text, cls) -> ?instance of cls */
/* `m[k]` on a map is the OPTIONAL read: a missing key is nil, not a trap.
* `get(m, k)` (WO_B_MAP_GET) stays the asserting form. Indexing a `multi`
* out of range still traps — a bad index is a fault, not an absence. */
WO_B_MAP_GET_OPT = 59, /* (map, key) -> value or nil */
};
#define WO_B_MAX 58u
#define WO_B_MAX 59u
/* ids at or above this one live in sysio.c, not builtin.c */
#define WO_B_SYS_FIRST WO_B_FS_EXISTS
/* json.encode's "this is already JSON" kind (a `json.Value`): not a field
* kind, only an argument marker. */
#define WO_JSON_KIND_RAW 255u
/* ---- instruction encode/decode: op:8 A:8 then B:8 C:8 or Bx:16 ---- */
static inline uint32_t wo_ins_abc(uint8_t op, uint8_t a, uint8_t b, uint8_t c) {
return (uint32_t)op | ((uint32_t)a << 8) | ((uint32_t)b << 16) | ((uint32_t)c << 24);

View file

@ -3,6 +3,13 @@
-- missing key traps KEY"). Unprovable at compile time -- the key is a
-- runtime `Text` value -- so it traps rather than failing to compile:
-- the hybrid boundary's runtime half.
--
-- Written as `get(...)` on purpose: the two map reads differ. `get(m, k)`
-- ASSERTS the key is there and traps KEY when it isn't (this fixture);
-- `m[k]` is the OPTIONAL read and yields nil for a missing key, which is
-- what makes `let v = m[k]; if v != nil { ... }` the ordinary lookup idiom
-- (the driving workload's own header lookup). Before 2026-08-14 both
-- spellings trapped, and this fixture used the sugar.
class Catalog {
prices: map<Text, Int>
}
@ -10,5 +17,5 @@ class Catalog {
fn main() {
let c = Catalog { prices: map_new() }
set(c.prices, "mug", 499)
print_int(c.prices["shirt"])
print_int(get(c.prices, "shirt"))
}