fix(compiler): apply OverPy's per-parameter False/True substitutions under optimizeForSize

[e54-bot]

Fixes #388

Summary

• Adopt workshop-rs 0.8.0, which stops normalizing contextual literals (fix(parser)!: preserve contextual literal substitutions workshop-rs#300).
• Under #!optimizeForSize, spell 0/1/Null/zero vectors as OverPy 9.7.10 does per parameter, from a table generated by tools/overpy/gen_literal_flags.cjs (OverPy's canReplace0ByFalse, canReplace1ByTrue, canReplace0ByNull, canReplaceNullVectorByNull). Policy stays in opy-rs and no longer reads catalog paramCoercions.
• Remove the unconditional lowering normalization (0→Null, False→0, empty array→"") that only converged because the old parser did the same. Unoptimized output keeps the authored spelling, as the pinned reference does (checked by compiling probes with pinned OverPy).
• Update tests that asserted the old normalized text to the oracle's output.

Verification

• cargo test --workspace, cargo clippy --workspace --all-targets, cargo fmt --check, tools/overpy/tests: pass.
• tools/overpy/probe_builtins.py: 4064 probes, unexplained: 0 (two approved gaps unchanged).
• tools/overpy/structural_gate.py: Bastion main.opy and externalMain.opy report 0 differences against pinned OverPy.
• Wright canonical diff (issue Apply OverPy's per-parameter False/True substitutions under optimizeForSize #388 AC): a clean Wright main worktree with only workshop-rs bumped to =0.8.0 (its main pins 0.7.1, which predates fix(parser)!: preserve contextual literal substitutions workshop-rs#300), compiling Bastion (f693704) through this branch's opy-provider, compared with workshop_rs::roundtrip::equivalent against pinned OverPy 9.7.10 output. main.opy (zh-CN): equivalent. externalMain.opy (en-US): equivalent, also via wright semantic-compare. wright semantic-compare only parses en-US, so the zh-CN comparison used workshop-rs directly with the zh-CN locale.

[Teakowa]

Three blocking findings:

1. crates/opy-rs/src/compiler/size_optimization/literal_flags.rs / tools/overpy/gen_literal_flags.cjs: the generated upstream canReplace* table is interpreted directly by generic Rust to choose observable OverPy lowering. That conflicts with the current language-core.md / AGENTS.md contract: source-language behavior and coercion/lowering policy belong in discoverable typed Rust; generated/declarative data must not become a semantic control layer, and upstream implementation data should not be mechanically imported into the core. Encode this policy directly in opy-rs typed code and use the upstream table/probe for completeness verification, or obtain an explicit owner architecture decision changing that contract.

2. Required verification is missing. AGENTS.md says lowering/optimization/builtin-argument changes must run tools/overpy/probe_builtins.py. The PR verification does not report it, and the current CI run does not execute that probe. Run the full builtin-call probe and resolve any unexplained differences.

3. Issue #388 explicitly requires Wright's canonical diff for Bastion main.opy and externalMain.opy; the PR says it was not run. The opy-rs structural gate is green for those same pinned entries, but review should not silently substitute a different check for an explicit AC. Run the Wright check, or have the owner update the AC if the structural gate is now the intended replacement.

[Teakowa]

Findings 1 and 2 are resolved. The generated upstream flags are now verification data only, while the policy is encoded in typed Rust, and the follow-up commit records the builtin probe with 0 unexplained differences.

One blocker remains: Issue #388 still explicitly requires Wright's canonical diff for Bastion main.opy and externalMain.opy, while the PR still says that check was not run. Run that acceptance check, or update the Issue contract if the opy-rs structural gate is now the intended replacement. No new code-level findings in the follow-up patch.

[e54-bot]

Thanks. All three findings are addressed in 5eb8b5c and the PR description.

1. The policy is typed Rust in literal_slots.rs. The upstream flags are verification data only (tools/overpy/upstream-literal-flags.json), checked in both directions by slots_cover_the_upstream_flags.
2. tools/overpy/probe_builtins.py: 4064 probes, unexplained: 0.
3. Wright canonical diff, run on a clean Wright main with only workshop-rs bumped to =0.8.0 (its main pins 0.7.1, which predates fix(parser)!: preserve contextual literal substitutions workshop-rs#300), through this branch's opy-provider: Bastion main.opy (zh-CN) and externalMain.opy (en-US) are both equivalent to pinned OverPy 9.7.10. Details are in the PR description. The Wright workshop-rs bump is being opened as a separate PR.

Requesting re-review.

[Teakowa]

LGTM. The three prior review findings are resolved, the Wright acceptance check is now documented for both Bastion entries, and the current CI run is fully green.