P6.x: correct the ast_importers completion criterion to its reachable floor

[accuser]

Summary

#1187 slice 6 — correcting the track's own stated completion criterion (ast_importers = 0), which #1187's own scoping work found was never actually reachable for three files.

AST_IMPORTER_EXCEPTIONS now names emitter.rs, emitter/lower.rs, and project/tests_emit.rs, alongside the existing ir.rs/ir/lower.rs pair.

emitter.rs/emitter/lower.rs: design/tracks/the-ir.md's own Q7 (§3.7) already settled that these files keep hand-writing TypeScript source text after this track's cutover — the printer that would let them stop reading raw Handler/ServiceDecl/AgentDecl parameters for header/body-rendering surface is phase 7's, explicitly out of this track's own scope. ast_importers = 0 was therefore never reachable for either file — the same structural reason it's unreachable for ir.rs/ir/lower.rs, just the opposite half of the Ast ⇄ Ir boundary (a lowering pass whose whole job is reading the AST, vs. a still-string-writing pass whose whole remaining job, past this track's scope, is reading it too).

project/tests_emit.rs was deliberately left out of the original exclusion — a test (ast_importer_exclusion_is_named_not_prefixed) used to assert exactly that, on the reasoning that its AST coupling was still-open, real work. This PR corrects that on new evidence, not a silent reversal: its test/suite case bodies call emitter::lower_block_to_async_body/lower_test_case_body/lower_integration_case_body directly (the same Q7-settled body-rendering pass), and its own driver_param_ty/strip_effect_httpresult read a handler's declared param/return TypeRef with no corresponding TyId available at that call site — the same caller-reads-callee's-declared-shape pattern #661 established for cross-context codec generation. project.rs, which prompted the original "must stay counted" stance, shares neither reason and stays counted.

ast_importers moves from its stale committed 8 to its corrected, reachable floor of 5 (runtime_use.rs, emitter/workers.rs, emitter/serialisation.rs, emitter/workers_entry.rs, project.rs remain — all genuinely open R6.13 surface).

design/tracks/the-ir.md §5/§6 record the correction and the full #1187 slice history since the doc was last updated, including slice 4 (serialisation.rs's codec-name mangling to TyId), which was scoped in full this session and then explicitly abandoned rather than landed — both of its headline claims (ast_importers 8→6, and a straightforward TypeBody→TypeShape conversion) turned out false once actually investigated. Recorded rather than silently dropped, per the track's own "the evidence ages" discipline.

Test plan

• cargo build --workspace
• cargo clippy --workspace --all-targets
• cargo fmt --all -- --check
• cargo test -p xtask --lib — all 99 tests pass, including the three ast_importers/exclusion tests updated to match the new list
• cargo test --workspace — full suite
• cargo xtask greenfield-status --apply — design/greenfield-status.md updated to the new committed value (8 → 5)

🤖 Generated with Claude Code

[claude]

Reviewed the four-file diff. Release discipline is right: exactly one design/pending/p6-ast-importer-exceptions.md with a level and a changelog, and the change is real code (a gated probe's counted value), so an increment is correct. The arithmetic checks out — 10 files in bynk-emit/src match bynk_syntax::ast, minus the five named exclusions leaves the committed 5, and the five survivors are exactly the ones named in the PR body. project/tests_emit.rs's inclusion is well-evidenced.

Three things I'd want resolved before this lands, in priority order.

1. The Q7 justification for emitter.rs/emitter/lower.rs doesn't match what those files actually import the AST for

Q7 (§3.7, as quoted in the-ir.md §5) settles that the string-writing shape survives the cutover — and says the opposite about the imports:

It means the AST-walking decisions those string-writing functions currently re-derive (Callee dispatch, commit shape, store-write target) get replaced by reads off an already-lowered IrExpr/IrItem/CommitShape/StoreFieldIr value; the functions' own Lowered-returning, string-writing shape survives the cutover unchanged.

The live non-test bynk_syntax::ast uses in these two files are the decisions, not header/body-rendering parameters:

• bynk-emit/src/emitter.rs:1400-1407 — walks commons.commons.items for CommonsItem::Service, then reads svc.protocol / ServiceProtocol::Events. That is verbatim what §6's P6.6 row is for ("closes the emitter reading AST declarations directly"), and P6.6 is in scope for this track.
• bynk-emit/src/emitter/lower.rs:1147-1153 (cap_op_param_names) — walks commons.items for CommonsItem::Capability. Same AST-declaration read.
• bynk-emit/src/emitter/lower.rs:4120 / :4172 — bynk_syntax::ast::queue_variant(&id.name), a name-matched variant dispatch: the exact class P6.2/Callee exists to replace, and a sibling of the block_writes_state name-matched-receiver defect §9 names as this track's live data-loss risk.

None of those need a phase-7 printer to go away; they need P6.2/P6.6, both in this track's own §6 table. Excluding these two files (10.7k lines, the two largest AST-coupled emitter files) means P6.2/P6.5/P6.6 can land in full and move ast_importers by exactly zero — the probe stops measuring the track's headline deliverable, which is the failure mode #1176's own named-not-prefixed argument was written to avoid.

The project/tests_emit.rs half stands on its own evidence (driver_param_ty/strip_effect_httpresult with no TyId at the call site) and looks sound. Suggest landing that one and keeping emitter.rs/emitter/lower.rs counted until the decision-walking above is actually gone — or, if the exclusion is kept, narrowing the stated justification to the surface Q7 really settles and saying plainly that the probe no longer covers P6.2/P6.6.

2. The doc now says two incompatible things about the criterion

design/tracks/the-ir.md:621 calls 5 "the completion probe's own true floor, not a residual number expected to keep falling." But four statements of the criterion are unchanged by this PR:

• §5:310 — "Here: ast_importers reads 0"
• §6:415 — "Completion probe: ast_importers = 0 (§5) — live today at 9" (also now stale: live is 5, not 9)
• §7:632 — phase 7's entry condition, "this track's probe (ast_importers) reads 0 (§5)"
• §12:715 — "retires when §6's probe (ast_importers) reads its named floor (0, per §5)"

If 5 really is a floor, all four are unsatisfiable and — on this doc's own terms — the track can never retire and phase 7 can never open. Either they move to the new floor, or the claim softens to "reads 5 today." The floor claim also sits badly with the sentence right before it: project.rs is listed as remaining "genuinely open R6.13 surface" and is P6.6's own named target ("EmitProjectCtx's 28 fields"), an in-scope slice. A number with an in-scope slice still pointed at it isn't a floor.

3. ast_importer_exclusion_is_named_not_prefixed no longer tests what it says it does

Both emitter/-prefixed negatives (emitter/lower.rs, project/tests_emit.rs) moved to the positive side, leaving project.rs — which isn't under emitter/ — as the only negative. An emitter/** prefix rule would now pass this test unchanged while wrongly excluding emitter/runtime_use.rs, emitter/workers.rs, emitter/serialisation.rs, emitter/workers_entry.rs. One line restores the invariant:

assert!(!is_named_ast_importer(Path::new("emitter/workers.rs")));

Same gap in ast_importers_excludes_the_named_pairs_but_counts_project_rs — assert!(counted.contains("emitter/workers.rs")) keeps the live-tree survivor-set assertion honest.

---

The honesty of the record is good — writing up abandoned slice 4 rather than dropping it is the right call, and the project/tests_emit.rs correction is properly argued as new evidence rather than a reversal. My concern is narrower: the emitter.rs/emitter/lower.rs half of the exclusion buys a lower number by removing the probe's coverage of the slices that haven't landed yet.

🤖 Generated with Claude Code