[M11 correctness] Close supported-surface Workshop emission and round-trip divergences

[Teakowa]

Parent: #82
Depends on: #86

Goal

Fix the class-3 emission divergence: string values inside Array(...) initializers must be emitted in the pinned oracle's Custom String("...") form, so supported-surface programs with string arrays produce N-level-equal Workshop output.

Context (QA class-3 finding, first of M11)

globalvar x = ["a","b"] emits native Array("a","b") vs oracle Array(Custom String("a"), Custom String("b")). Corpus instance: pixelart.opy:35 globalvar owo = [' ', ...] diverges at normalized char 411; pixelart full compile exit 0 and settings-section equality hold, but full-program normalized equality fails. Arrays + strings are inside the declared supported surface (support-matrix); no documented intent (not class 5); complete evidence (not class 6). Pre-existing M8 emitter bug, surfaced by the first full-program compile; not a settings regression. The 6 v1-gates fixtures pass because none contains a string array.

Scope (bounded emitter correctness fix)

• crates/wright-workshop/src/emitter.rs: render string values in the contexts the pinned oracle wraps with Custom String("...") — known instance: direct string elements of Value::Array (the shared native/adapter emission path covers both). Match the oracle's string-wrapping in value positions for every context evidenced by the minimal repro and the corpus snapshots; no context that currently matches the oracle may change.
• Regression tests: minimal repro (globalvar x = ["a","b"] + one rule) byte-asserted Array(Custom String("a"), Custom String("b")); pixelart full-program normalized-equality check (v1 normalizer: debug-HUD collapse + whitespace collapse); the 6 v1-gates rows still green; the 7 settings sections still equal (pixelart 256, santa 350, broken-weapons 511, client-to-server 297, parabola 136, crosshair 144, inputhud 476 — whitespace-collapsed); ws-parser roundtrip for the new spelling.

Non-goals (hard)

• No deferred syntax (settings semantics, \, ++/--, dicts, """, #!obfuscate, expression-level in/not in, named arguments, range literals, Team.2, inline conditionals, comprehensions, bytes literals, multiline strings).
• No normalizer change to mask the divergence (would manufacture a class 5 without documented intent).
• No v1-gates FIXTURES change. A pixelart gate row may be proposed later only after verification, with inventory evidence.

Acceptance criteria (QA-verifiable)

• AC-1: minimal repro emits Array(Custom String("a"), Custom String("b")) (byte-asserted test); pinned-oracle artifact identical under the v1 normalizer.
• AC-2: pixelart full-program normalized equality (normalizedEqual: True in target/m11-nlevel.json, nativeExit 0) — or, if any divergence remains, QA records the precise remaining divergence list and reclassifies each.
• AC-3: previously-passing emission contexts unchanged: v1-gates 6/6 (FIXTURES unchanged); the 7 settings sections still equal; roundtrip/emitter suites green.
• AC-4: no new class-3 findings: per-program first-failure matrix unchanged or strictly improved (pixelart becomes the first full-compile + full-equality program); oracle 20/20; adapter 22/22; differential green (PARITY_CASES 7 rows; a pixelart row only if native==adapter HIR subtree equality is separately verified — never forced); all cargo suites; six CI jobs green.
• AC-5: inventory refreshed at the fix commit (class-3 resolved; parity baseline; pixelart row status exactly as verified).

Admin note

#86 closes with AC-3 recorded as contingent-not-justified (its acceptance record is complete); this issue carries the pixelart full-program row forward (N-level AC; PARITY_CASES row contingent on a new native==adapter HIR assertion).

PM amendment — extended acceptance criteria (2026-08-14)

QA verification at dbd342e (docs/opy/m11-issue87-verification.md) found AC-2 not met: pixelart full-program normalized equality fails (279 regions / 10,481 positions, all in the 9 artwork string-array sites) with three recorded mechanisms. The residual set is the same emission contract as this issue; the acceptance criteria are extended within #87 (no new issue).

• AC-1 (stand): minimal repro Array(Custom String("a"), Custom String("b")) byte-equal; the test artifact uses a real action body (no pass), oracle-identical end to end.
• AC-2 (amended): pixelart full-program normalized equality (normalizedEqual: True in target/m11-nlevel.json) OR a residual list shrinking to exactly what remains after A/B/C, each item reclassified; mechanisms A and B must be gone.
• AC-3 (new): long-string split parity — oracle rule: non-final segments = exactly 125 content chars + {0} (128 total), chained Custom String("...{0}", Custom String("...")), final segment = remainder; synthetic 300-char (3 segments) and 1000-char (8 segments) repros byte-equal to the pinned oracle artifact (byte-quoted in tests); char-count basis pinned from the oracle artifact.
• AC-4 (new): value-string re-escape parity — globalvar x = "a\nb" emits Custom String("a\nb") with literal \n (0x5C 0x6E), byte-equal to oracle; \ " \t \r handled per oracle evidence; the 7 settings sections still equal (256/350/511/297/136/144/476) proving the settings path untouched.
• AC-5 (new): empty-rule drop parity — pass-only rule and condition-without-actions repros emit no rule; artifact byte-equal to the oracle.
• AC-6 (new): roundtrip fixed-point for the new spellings — split strings and re-escaped strings parse and re-emit byte-identically; a bounded ws-parser gain for the oracle's Custom String chain shape is authorized if the generic call path cannot consume it (self-contained; no general call rework).
• AC-7 (stand): previously-passing contexts unchanged — v1-gates 6/6 (FIXTURES unchanged), emitter/driver/roundtrip suites green.
• AC-8 (stand): 12-program first-failure matrix unchanged; no new class-3; oracle 20/20; adapter 22/22; differential green (PARITY_CASES 7; a pixelart row only with a committed native==adapter HIR subtree assertion — never forced); all cargo suites; six CI jobs green.
• AC-9 (stand): pixelart parity-row decision per AC-8; N-level row status exactly as verified.
• AC-10 (new): inventory refresh at the fix commit (class-3 A/B/C resolved; pixelart row status; supersede the 4f41cdf residual sections where they differ).

Non-negotiables: no normalizer masking; no v1-gates FIXTURES change; no deferred syntax; the settings emission path untouched.

PM amendment 2 — final correctness batch (AC-11..AC-14, 2026-08-14)

QA's pre-gate scan (docs/opy/m11-inventory-final.md, a87b0ec) resolved the pixelart class-3 family and surfaced three further class-3 findings on supported surface, all repro-pinned: (1) trailing-if End; omission + ws-parser roundtrip asymmetry; (2) .format() constant folding (semantically equivalent); (3) non-default numeric globalvar initializers dropped (semantic — variables start at 0; live adapter carries them; the support-matrix claim is refuted and corrected in-batch). All three are authorized as one bounded batch inside #87.

• AC-11 (numeric initializers): j = 5 preserved in HIR and emitted Set Global Variable(j, 5) in the Initialize rule byte-equal to oracle; h = 0 dropped (matching oracle and adapter); k = 0.0 preserved with the oracle's spelling; bare-index form untouched; new synthetic fixture (e.g. synthetic/declarations-numbers) with adapter fixture + oracle.json + PARITY_CASES row (HIR parity) + N-level; support-matrix claim corrected: "integer-0 literal initializers are dropped from HIR (matching the reference adapter); non-zero and non-integer numeric initializers are preserved".
• AC-12 (trailing-if): rule-final if/if-else emitted without the closing End; byte-equal to oracle (repro pins a-d: rule-final if, rule-final if-else, if inside rule-final for/while, middle-of-rule if which keeps End); ws parser accepts the oracle's spelling (bounded, rule-block-close-only); roundtrip fixed-point green.
• AC-13 (format folding): all-constant numeric-arg .format() folds to Custom String("value: 3") byte-equal (repros: single arg, multi-arg, float spelling 0.5); variable-arg case unchanged byte-equal; HUD/debug path untouched; folded output feeds splitting/re-escaping.
• AC-14 (no regression): v1-gates 6/6 (FIXTURES unchanged); 7 settings sections equal; pixelart N-level still True; 12-program first-failure matrix unchanged; differential green (PARITY_CASES 7 or 8 per AC-11); oracle 20/20; adapter 22/22; all cargo suites; clippy 0; fmt clean; six CI jobs green at the batch commit; inventory refreshed (superseding a87b0ec's findings sections); no new class-3 anywhere.

Non-negotiables: no deferred syntax; no normalizer masking; no v1-gates FIXTURES change; no settings-surface change; no E-level work. After the checklist passes, the #82 final gate runs (accept-with-limitations criteria per the #82 decision record).

PM amendment 3 — final batch (AC-15..AC-17, 2026-08-14)

QA's final checklist (ea2c1a4) verified the full batch but the final scan found two more class-3 items, both emission/roundtrip canonicalization on matrix-listed supported surface (corpus-unexercised, zero first-failure/parity impact): (1) format placeholder canonicalization — "v: {}".format(x) native emits {}, oracle emits {0}; (2) playervar member reads in value positions — native Event Player.p, oracle (Event Player).p, and both spellings fail the native ws parser (native emission does not round-trip). Authorized as the final batch:

• AC-15 (placeholder canonicalization): implicit {} rewrites to the oracle's explicit numbered form per pinned repros (single/multi/mixed explicit+implicit/repeated-arg shapes — pin from oracle artifacts FIRST); HIR format text stays {} in both producers (differential unaffected); constant-fold and HUD/debug paths untouched; un-pinned shapes unchanged with a fixed-point parseability assertion.
• AC-16 (playervar reads): g = eventPlayer.p emits Set Global Variable(g, (Event Player).p) byte-equal to oracle; ws parser accepts the oracle's spelling (bounded); roundtrip fixed-point green; the SET form and method-call receivers untouched.
• AC-17 (closure scan + no new class-3): the ENUMERATED matrix-surface closure scan runs clean — every emitter-produced spelling self-roundtrips (parse(emit(x)) byte-identical) and matches the oracle's canonical form where a repro compiles; 12-program first-failure matrix unchanged; v1-gates 6/6 (FIXTURES unchanged); 7 settings sections equal; pixelart normalizedEqual True (19,925/19,925); PARITY_CASES 8 unchanged; oracle 21/21; adapter 23/23; all cargo suites; clippy 0; fmt clean; six CI jobs green at the batch commit; inventory refreshed (superseding ea2c1a4's open-items sections).

Scan boundary (decision): the no-open-class-3 criterion is scoped to the enumerated matrix-listed emission surface (self-roundtrip closure + oracle-canonical repros); constructs outside the matrix are deferred by definition; this is the LAST scan — no further ad-hoc scans. After (a)+(b) and a clean closure scan, the #82 final gate runs as accept-with-limitations per the #82 decision record.

PM amendment 4 — final single item (AC-18, 2026-08-14)

The closure scan's sole remaining finding (augmented playervar assignment eventPlayer.p += 1: native Set Player Variable(..., Add((Event Player).p, 1)) vs oracle Modify Player Variable(Event Player, p, Add, 1); matrix-listed, oracle success, semantically equivalent, both spellings roundtrip, zero corpus coverage) is accepted as class 3 and authorized as the final single-item batch:

• AC-18: pin the oracle's per-operator Modify spellings for playervar augmented assignments from byte-quoted artifacts first (+= -= *= and any the oracle renders as Modify Player Variable; operators the oracle does NOT support via Modify — per actual behavior e.g. /= //= %= if un-evidenced — stay on the current Set (...) path with a recorded note); fix the HIR->WIR lowering so playervar augmented assignments produce ModifyPlayerVariable for the oracle-evidenced op set, mirroring the already-byte-equal global path (shared by native and adapter-driven lowering; HIR unchanged; differential/PARITY_CASES 8 unaffected); add the playervar augmented-assignment family to the closure scan (roundtrip fixed-point). Regression: global g += 1 unchanged; pixelart normalizedEqual True; 7 settings sections; v1-gates 6/6; PARITY_CASES 8; 12-program matrix; oracle 21/21; adapter 23/23; all suites; clippy/fmt; six CI jobs green; inventory refreshed. The eventPlayer.getCurrentHero() probe is closed as outside-surface (catalog-data gap, unknown-enum-member precedent — deferred data change, not class 3).

[Teakowa]

Closing #87. Full closure record AC-1..AC-18, independently verified (m11-issue87-verification.md, records 4f41cdf/ea2c1a4/556e9fc/b9a2309; CI 31770753398, 31772287928, 31773787648): the complete supported-surface emission class-3 family is resolved with repro-pinned oracle parity — Custom String wrapping (dbd342e), long-string splitting 125+{0}/128 per segment (b5b0578), value-string re-escaping (b5b0578), empty-rule drop (b5b0578), numeric initializers preserved except integer-0 + support-matrix correction + synthetic/declarations-numbers (f71bc4a/c8e3430/8182959), trailing-if End; + ws-parser roundtrip symmetry (1841452), format constant folding (1841452), {0} placeholder canonicalization + playervar-read (Event Player).p spelling (65b7fb6), augmented playervar Modify Player Variable spelling — 5 operators byte-equal, //= rejected by both engines (doc-accuracy note only; da38f57). The enumerated matrix-surface closure scan (19 families) is final-clean: no in-surface class-3 remains. Pixelart full-program N-level row normalizedEqual True (19,925/19,925); PARITY_CASES 8; v1-gates 6/6 unchanged; oracle 21/21; adapter 23/23; 12-program matrix unchanged; all suites green at da38f57/b9a2309. HIR and the differential stayed untouched throughout (lowering-only and emitter-only changes). AC-8/AC-9 honored: no parity count forced.