Model census heads in the authority and drop non-closure bodies after parsing

[gunbai-bot]

Entry-scoped compilation still retained parsed bodies for every non-closure census module after #11072 streamed their tokens. This change models the declaration-head projection in v1.compiler.compile, then retains that projection after full parsing and annotation admission. The seed delegates to the generated projection instead of maintaining a second strip implementation.

The projection preserves declaration identity, names, kinds, signatures, imports, and spans. Functions retain the parser’s refusing body marker for function-kind discrimination; executable bodies, uses, match patterns, and expression data are discarded. Both census accumulator references share the projected parse result. Full-body parse errors still reach the compile refusal.

Only compile_to_resolved_with_options selects RetainDeclarationHeads for its non-closure sources. The existing parse_census_fill_sources API retains full nodes for namespace-wave admission, singleton CLI admission, entry resolution, and the diagnostics-only compile-clean caller. This avoids changing occurrence/namespace consumers for no meaningful singleton retention saving. Closure modules continue through the normal frontend. Frontier (b), the whole-pool edge index, is separate.

Validation:

• Native arm64 release build (-p v1-compiler --bin gunbc --bin claim_executor -j 2) and cargo fmt --all --check pass. Binaries were copied out of shared /cargo-target immediately after each build.
• Model-first commit f6581910ab2 contains the authority projection and its generated mirror. Retention and seed delegation follow in 812c542d491.
• claim_executor --required-regen --source-root dag --source-root src/v2 passes at 812c542d491: first_generation_equal=true, planned/executed/adjudicated 155/155/155, changed paths empty. Authority digest fnv1a64:1321de6fd99aff1b; committed/candidate generated digest fnv1a64:d70e079d37a101da.
• Only the emitted v1_compiler_compile.rs candidate was installed, using the parent-approved interim route for the installer gap tracked by Include host-shell mirror ownership in regen installation #11094. No generated Rust was hand-edited.
• Rebuilt-seed claim_executor --required-regen-fixed-point --source-root dag --source-root src/v2 passes: fixed_point_equal=true, referenced_first_generation_equal=true, both at 812c542d491. CI build (including cargo clippy --all-targets -- -D warnings) and generated-artifact regeneration pass. Dashboard review 64084 approves with no findings. The witness lane is infrastructure-refused; see below.

Isolated before/after, same worktree corpus and invocation (16 closure sources, 5,429 census modules):

Run	VmHWM (KiB)	GiB	Wall (s)	Exit	Margin to 15 GiB
Streaming baseline, positive	9,063,388	8.644	177.530	0	6.356 GiB
Heads retained, positive	7,886,192	7.521	167.962	0	7.479 GiB
Streaming baseline, balanced-body red	9,063,308	8.643	161.038	1	6.357 GiB
Heads retained, balanced-body red	7,886,224	7.521	169.985	1	7.479 GiB

The measured positive saving is 1,177,196 KiB = 1.123 GiB (12.99%). This is smaller than bright-bear-236’s prior heads-retention probe (8,986,140 → 7,663,268 KiB, about 1.26 GiB); #11072’s streaming-only result was 8,986,500 KiB after 10,006,804 KiB. Those are predecessor measurements, not substituted for this session’s results. The admitted budget stays 16,106,127,360 bytes (15 GiB). No floor budget rows or parse witnesses changed, and no floor JOB verdict is claimed as proof.

Both positive runs report 0 blocking errors, 213 advisories, and one emitted file. The discriminating red appends fn body_probe(x: Int) -> Int { let = = = 9 } to non-closure dag/std/access.dag, then restores the file. Both runs refuse at emit with two hard diagnostics and the identical error[dag/std/access.dag]: expected name, found Eq (span 2646–2647). Balanced braces make this invisible to the heads index; the full-body census parse must catch it. Artifact byte equality is not used because registry ordering is nondeterministic on the baseline.

Reproduction (run the identical command against each snapshotted binary, polling /proc/<pid>/status for VmHWM and timing with a monotonic clock):

gunbc compile --source-root dag --source-root src/v2 \
  --entry src/v2/workflow/floor2_prepared_subject.dag \
  --output-dir /tmp/heads-measure-output --target dag

Baseline gunbc SHA-256: c5c07879131d35731fec3f90c1524d8acdf128e9f8b9c13fae75142c33e918b3.
Final gunbc SHA-256: d0a38657f28295aaacdcba875557617896627741211dd45b11a305bce8edb375.

CI disposition: witness job 103431616486 refused with MemoryStallRefusedPageThrash while resolving floor_terminal_ledger_wire and typechecking floor2_prepared_subject: 282,224 major faults/minute, 4% CPU share over 34,140 ms, admitted budget unchanged at 15 GiB. The terminal ledger could not be rendered, so this is explicitly not a green floor. The run classified itself class=infra; it is a draw, and the isolated before/after measurements above remain the evidence for this change. GitHub is blocked pending disposition of that check.

[gunbai-bot]

Operator handoff for 812c542d491, under the parent’s refused-floor disposition:

1. Build and generated artifacts are green. Required build, cargo clippy --all-targets -- -D warnings, and heal-generated-artifacts pass. Local first-generation equality covers all 155 mirrors; the rebuilt-seed fixed point also passes at this head.
2. The floor is a draw, not green. Run 34649265163, witness job 103431616486 refused with MemoryStallRefusedPageThrash, explicitly floor_class=infra, during terminal-ledger grammar resolution. No rerun or budget increase is being used to manufacture a pass.
3. The isolated retention saving is 1.123 GiB. The same entry invocation and 16-source/5,429-module population went from 9,063,388 to 7,886,192 KiB (8.644 → 7.521 GiB), with positive wall time 177.530 → 167.962 seconds. Margin to the unchanged 15 GiB budget is 7.479 GiB. The PR body carries commands and binary hashes.
4. The discriminating parse red is preserved. The balanced-brace non-closure body probe refuses before and after with exit 1, two hard diagnostics, and error[dag/std/access.dag]: expected name, found Eq at the same 2646–2647 span. Bodies are fully parsed before projection; the probe was restored.
5. Review is approved. Dashboard review 64084 (cursor) approves with no findings; its complete artifact was read and there are no outstanding review changes.

Implementation and evidence are complete. Per sunny-tern-813’s ruling, stopping here for the operator to land this PR; GitHub’s refused-floor check remains explicitly red.

— sent from valiant-badger-803