W7.3 grammar-expansion (follow-up к #46)

[gHashTag]

W7.3 grammar-expansion (продолжение PR #46)

Base: main @ 3272583 (PR #46 merged, E1+E2 baseline frozen — 100/100/0 subset)
Head: 9bbc103
Scope: grammar-expansion до E3 unblock (E3 timer-blocked до 2026-07-19 12:24 UTC per PR #44 backstop)

Commits (по SHA, chronological)

2080510 — \b-fix normalize_ast (carry-over из GLM PR #46 review)

Non-blocking precision-nit из GLM peer-review PR #46 @ 4a016d4:

• Regex line:\s*\d+,? без \b теоретически over-match на *line: N полях
• Empirical verification на N=1000 frozen baseline — nit latent, не active
• Fix defensive против future parser evolution
• Regression check N=50 subset → 50/50/0 ok

247427d — grammar-expansion target #1 (collection-typed params) — OBSOLETED by 9bbc103

Emitted Zig-style []const T / []T / [N]T collection forms. GLM peer-review flagged blocking finding: syntax mismatch with the actual W6.2 audit corpus at tri-net/specs on PR #39 branch (that corpus uses Rust-style [T; NAMED_CONST] exclusively). Retained in history for traceability; replaced by 9bbc103.

9bbc103 — collection-params rewrite to [u32; NAMED_CONST] (audit-corpus-aligned)

Corpus-mismatch resolution:

The audit corpus targeted by W6.2 lives at tri-net/specs on the PR #39 branch (feat/strategic-audit-2026-07-04, 68 spec files). Verification against that corpus:

• 0 []const T occurrences (Zig-style is absent).
• 159 [u32; NAMED_CONST] occurrences across the 68 files.
• 100% u32 element type.
• Top declared consts by count: MAX_NODES (29), MAX_PARAMS (18), MAX_METRICS (12), MAX_FLOWS (11), and so on.
• Const-decl syntax: const NAME: u32 = <literal>; at module scope, literals in [2, 32].

247427d was chosen from a workspace-wide grep that also matched a parallel non-tri-net corpus at ../t27/specs/ (830 []const T occurrences there). Corpus was not scoped to tri-net/specs. Anchor recorded (Anchor #4): any claim verified against ground truth requires scoping the verification tool to the same corpus as the claim.

Revision applied (option a — drop Zig-forms + u32-only):

• pick_coll_type rewritten to emit only [u32; NAMED_CONST].
• Ctx.declared_consts: Vec<(String, u32)> added for per-module tracking.
• gen_const_decls: 1–4 module-scope const NAME: u32 = <literal>; per module, values in [2, 32], no repeats within a module.
• NAMED_CONST_POOL = top-10 audit-corpus names (MAX_NODES, MAX_PARAMS, MAX_METRICS, MAX_FLOWS, MAX_ENTRIES, MAX_MODULES, MAX_TASKS, MAX_FUNCTIONS, MAX_SAMPLES, MAX_RESULTS).
• gen_module emits const-decls before fn-decls.
• gen_params guards: emits collection-param only if declared_consts non-empty, else falls back to scalar.
• Isolation constraint preserved: collection-typed idents recorded in Ctx.coll_params, never pushed into Ctx.idents.
• Grammar header comment updated.

Re-baseline N=1000 (this HEAD 9bbc103):

Metric	Frozen subset (3272583)	Expanded (9bbc103)
Parse-success	1000/1000 (100.0%)	1000/1000 (100.0%)
Determinism	1000/1000 (100.0%)	1000/1000 (100.0%)
Whitespace-invariance	1000/1000 (100.0%)	1000/1000 (100.0%)
Parse errors / Panics / Non-det	0 / 0 / 0	0 / 0 / 0
Fns emitted	~1900 (0-param only)	1951 (mix 0–4 params)
Fns with ≥1 collection-param	0	1177 (60.3%)
Fns with 0 params	~1900	392 (20.1%)
[u32; NAMED_CONST] occurrences	0	1924
Module-scope const-decls	0	2510 (avg 2.5 per module)

Sample module snippet:

module W73Fuzz0 {
    const MAX_SAMPLES: u32 = 25;
    const MAX_ENTRIES: u32 = 27;
    const MAX_TASKS: u32 = 31;
    const MAX_PARAMS: u32 = 18;

    fn f0(p_4412: u8, p_8371: [u32; MAX_TASKS], p_5676: [u32; MAX_SAMPLES], p_3484: u64) -> u16 {
        ...
    }
}

Планируемые следующие коммиты (tracked, не landed — separate PRs / commits per isolate-variables)

Priority по strategic value:

1. ✅ Collection-typed params [u32; NAMED_CONST] — LANDED @ 9bbc103 (superseded 247427d)
2. ⏳ Index expressions (arr[i]) — next; makes collection-params live in body (actual W6.2 Class 2 defect-surface exercise). Isolate-variables: alone, re-baseline, then next.
3. ⏳ Call — requires fn-signature registry in Ctx
4. ⏳ If-statements — Terminal-Return invariant extends to both branches

SHA-advance rule application

Per docs/W7_COLLAB_OPTIONS.md §SHA-advance rule (merged via PR #45 → main @ be06148):

• Prior review target: 247427d (Zig-style, BLOCKING FINDING from GLM re-review — syntax mismatch with tri-net/specs corpus).
• SHA advance: 247427d → 9bbc103 — delta:
  • - Zig-style pick_coll_type (three-form emitter for []const T / []T / [N]T)
  • + Rust-style pick_coll_type (one form: [u32; NAMED_CONST])
  • + Ctx.declared_consts: Vec<(String, u32)> field
  • + gen_const_decls — module-scope const-decl emitter (1–4 per module, [2,32] literals)
  • + NAMED_CONST_POOL — 10 audit-corpus top consts
  • + gen_module emits const-decls before fns
  • + gen_params guarded on declared_consts non-empty
  • Grammar comment updated (Module includes ConstDecl* prefix; CollType is [u32; IDENT])
  • ~ docs/W7_3_FUZZ_BASELINE.md §Expanded baseline rewritten (new §Corpus-mismatch resolution + new numbers table + new interpretation)
  • Isolation constraint retained unchanged
  • No changes to roundtrip.py

Reviewer: re-confirm against 9bbc103 per rule (Re-reviewed at 9bbc103: delta <bullet-list> in review comment).

External-dep window

• E3 unblock backstop: 2026-07-19 12:24 UTC (14 дней с PR W7.1 investigation — E0425 root cause is dropped let statements in t27c Rust emitter (findings-only) #44 approve @ 2026-07-05 12:24 UTC)
• Terminal-event overrides: t27#1401 won't-fix / closing PR / explicit reject
• Whichever comes first

Discipline compliance

• Draft-only per house rule (no merge until reviewed)
• Empirical grounding: every claim verified against fetched output on this commit
• Frozen citation-point PR feat(W7.3): grammar-directed fuzz baseline (E1 generator) #46 @ 3272583 preserved as reference
• Isolate-variables: only one grammar-target per commit (this = collection-params, revised for corpus alignment)
• Corpus-scope discipline: syntax choice verified against tri-net/specs on PR docs: strategic audit 2026-07-04 (7 findings + 3 options) #39 branch (the actual W6.2 audit corpus), not workspace-wide grep
• No pre-silicon claims (fuzz corpus is real t27c parse output)

phi^2 + phi^-2 = 3

---

Path-confirmation spot-check (GLM condition @ 9bbc103)

Peer-review of 9bbc103 approved the code and doc, gated merge on one path-confirmation check: verify that t27c gen-rust on a fuzz-generated spec produces Vec<> in param position (the W6.2 Class 2 defect surface). Executed 2026-07-05, full 1000-spec sweep:

Metric	Value
t27c gen-rust success	1000 / 1000
t27c gen-rust failed	0 / 1000
[u32; NAMED_CONST] in .t27 input	1924
Vec<> in gen-rust Rust output	1924
Match	exact one-to-one

Sample (spec fuzz_00000_seed_0x00c0ffee.t27):

Input (t27):

    const MAX_SAMPLES: u32 = 25;
    const MAX_TASKS: u32 = 31;
    fn f0(p_4412: u8, p_8371: [u32; MAX_TASKS], p_5676: [u32; MAX_SAMPLES], p_3484: u64) -> u16 {

Output (t27c gen-rust):

pub const MAX_SAMPLES: u32 = 25;
pub const MAX_TASKS: u32 = 31;
pub fn f0(p_4412: u8, p_8371: Vec<>, p_5676: Vec<>, p_3484: u64) -> u16 {

Confirms: syntax match ↔ codegen path match. Every collection-param position in the fuzz corpus exercises the W6.2 Class 2 defect surface (Vec<> param-position), not a divergent lowering path.

Doc-nit correction (GLM cosmetic)

Prior body said "pick_coll_type rewritten". Precise wording: the standalone pick_coll_type function from 247427d was removed; the equivalent collection-type emission is inlined into gen_params as format!("[u32; {}]", const_name) at gen.rs line 229. Semantics identical, structure simplified. Not a behavior delta.

phi^2 + phi^-2 = 3