[M11] Evidence-first real-world compatibility hardening (phase 1)

[Teakowa]

Goal

Build enough real-world compatibility evidence to determine what Wright actually needs next. Phase 1 is evidence acquisition and classification only; it does not implement newly discovered language features. The classified gap inventory decides which implementation issues should exist later. Completion of phase 1 is NOT M11 completion.

Phase-1 scope (SPEC-M11-phase1-real-world-corpus, owner: PM)

1. Define M11 requirements/spec for real-world corpus expansion.
2. Define the QA evidence/test plan.
3. Curate/acquire several independent, non-synthetic real-world .opy programs/projects with materially different structures.
4. Record provenance, licensing/redistribution status, acquisition method, source hash, pinned reference identity, and applicable compatibility gates.
5. Run the existing Wright/native-vs-OverPy compatibility machinery across the expanded inputs.
6. Produce a reviewable six-class gap inventory.

Requirements

• REQ-001: >=6 real-world fixtures with structural diversity, independent origins, a >100 KB fixture, and multi-file coverage.
• REQ-002: provenance/metadata completeness and reproducible acquisition/hashes.
• REQ-003: pinned OverPy oracle snapshots for committed fixtures.
• REQ-004: adapter HIR evidence or verbatim adapter failure for each fixture.
• REQ-005: native differential parity remains hard-fail for supported parity cases; gap fixtures are classified separately.
• REQ-006: per-fixture N-level inventory evidence without changing the frozen v1 release-gate fixture list.
• REQ-007: human + machine-readable QA-owned gap inventory with an explicit evidence snapshot.
• REQ-008: no newly discovered language feature implementation during phase 1.
• REQ-009: bounded post-M10 documentation drift corrected without behavior changes.

Accepted phase-1 findings

• Expanded corpus: 9 committed real-world fixtures plus 1 metadata-only record, 286 source files, 3 independent origins.
• 0/9 new fixtures are currently consumable end to end.
• Primary classifications: class 2 explicit unsupported surface ×5; class 4 reference/oracle limitation ×4; class 6 inconclusive ×1; no class 1/3/5 findings.
• Four fixtures are blocked at a top-of-file custom-game-settings boundary, so their program bodies remain unexercised by Wright.
• A real correctness defect was discovered: lex errors originating in an included file can be reported under the including file's path while retaining the included file's line/column.
• No phase-1 finding authorizes settings parsing, line-continuation support, postfix ++, dict literals, triple-quoted strings, or other language-surface expansion.

Evidence

• QA inventory: docs/opy/m11-gap-inventory.md and JSON companion.
• Inventory snapshot: 2ecc024; inventory commit: 394a2db.
• Reproducibility follow-ups through 5d234dc.
• Latest containing main CI: run 31728347878, all six jobs executed successfully.

Status

PHASE 1 COMPLETE. The evidence is sufficient to rebaseline M11. Follow-up implementation/evidence work belongs under the M11 parent created after this phase rather than expanding this issue.

[Teakowa]

M11 phase-1 — PM assessment (spec REQ-001..009 accepted)

Headline: Phase-1 deliverables are complete and REQ-001..009 are met. The expanded real-world corpus (9 committed fixtures + 1 metadata-only, 286 source files, 3 independent origins) is reproducible and provenance-complete (286/286 hashes, 3 upstream byte-checks). Machinery is green: oracle 16/16, adapter 21/21, differential OK with PARITY_CASES unchanged at 6, v1-gates 6/6 unchanged. Result: 0/9 fixtures consumable end to end — native 9/9 fail at the first out-of-surface construct, adapter 9/9 fail, 4/9 (pixelart, santa, broken-weapons, client-to-server) fail both engines at a top-of-file settings block, zero new parity rows, and zero class-3 findings (no supported-surface construct failed).

Inventory summary: class 2 (explicit unsupported surface) x5; class 4 (reference-oracle limitation) x4; class 6 (inconclusive/metadata-only, zombies) x1; no class 1, 3, or 5. Per-finding sub-rows in docs/opy/m11-gap-inventory.md (+ JSON companion), snapshot at 2ecc024, committed 394a2db.

Evidence boundary: S/D/N only — oracle overpy@9.7.10 pinned (gitHead 1e26889, metadata unchanged); native results at 2ecc024; adapter pinned with verbatim errors; per-fixture evidence boundaries are asymmetric (native stops at the first unsupported token; the oracle compiles full files or fails late) and are recorded per fixture. Oracle diagnostics are now checkout-path-independent (<fixture-dir> normalization, 5d234dc).

Real bug found (recorded, not misclassified): included-file lex-errors report the including file's path with the included file's line/col (repro: include a file containing a lex-error; e.g. 1v1_main.opy:7:66 vs real env.opy:7:66). Not class 3 by the strict definition (triggering construct is outside the supported surface; no oracle diagnostic-location parity), but a genuine diagnostic-source-attribution defect on a supported feature.

Minor mechanism notes: (1) the REQ-001 per-category cap was met in intent (diversity over count, all categories, 3 independent origins) but the matrix records membership, not primaries; (2) full-input-set hashes live in fixture.json files maps while oracle.json input.sha256 covers the main file only. Neither requires follow-up beyond record.

Smallest next batch (recommendation only — no implementation issues pre-created): one issue containing (a) the included-file diagnostic source-attribution fix (bug, correctness, D-level + LSP impact, with regression test) and (b) support-matrix deferred-section alignment naming \, ++, dict literals, """ with their observed rejection codes (contract fix, zero behavior). Explicitly deferred: (c) a structured unsupported-construct diagnostic for settings blocks — needs its own PM-specified diagnostic-only issue (frontend-change boundary), gated after (a)+(b); (d) a settings-free parity candidate — next phase, shortlist: skirmish_elim.opy, parabola.opy, crosshair.opy, inputhud.opy, lucioball_all_heroes.opy @ eea67ad, with a mandatory pinned-oracle pre-check before commit.

Governance note: the span-attribution defect is recorded in the inventory for Engineer confirmation; per the phase contract, no implementation issue is created from it (or from any class-2/4 finding) without PM prioritization, which this assessment now provides for (a)+(b). QA verifies the batch; Architect reviews the (b) contract edit. No phase-1 finding licenses feature work: settings parsing, \, ++, dicts, and """ remain deferred surface.