Fill composition overlays the closure onto the shared underlay (per-entry assembly)

[gunbai-bot]

Summary

Per-entry index assembly composed an entry's closure census with a shared per-MultiEntryIndex underlay memo — the whole-pool qualified fill, and the whole-tree bare census for each source root the entry reaches — by walking the underlay's key list, probing the accumulator per key, and path-copying an insert for every miss. The underlay is constant across entries while the closure varies, so every entry paid O(|whole pool|) work and retained O(|whole pool|) fresh HAMT nodes to build an index that differs only by its own closure.

Both directions denote the same map — union of the key sets, closure value wins the intersection — so overlaying the closure onto the shared memo is identical at O(|closure|), from an O(1) persistent clone.

Authority is src/v1/04_infer.dag (symbol_index_with_qualified_fill, symbol_index_with_bare_fill); v1_compiler_infer.rs is the regenerated seed. DESIGN §6 bare-minimum-cost: a copied accumulator is fixed regardless of the realized n, and the identity of the two directions is what makes this a cost fix rather than a semantic change.

Measured benefit

Fixed 50-entry cohort, one process holding one MultiEntryIndex, both arms built from one tree with one variable, two reproductions per arm:

	before (r1 / r2)	after (r1 / r2)	change
elapsed wall	268.4s / 248.0s	200.9s / 183.3s	−25.6%
peak RSS	7.78 GiB / 7.73 GiB	5.54 GiB / 5.53 GiB	−28.7%
additive resolve (50 resolves)	155.7s / 146.5s	105.0s / 94.6s	−33.9%
Σ exclusive assembly	100.8s / 84.0s	41.0s / 36.0s	−58.3%
symbol-index merge	34.7s / 28.4s	0.62s / 0.59s	−98.1%
per-root symbol-index composition	27.1s / 18.6s	0.72s / 0.62s	−97.1%

≈66 seconds recovered across 50 entries (1.32 s/entry in this cohort). Every aggregate is lower in both after runs than in both base runs, with no overlap. No fleet figure is claimed: these are additive spans plus this cohort's elapsed wall, not a required-path floor result.

Equivalence

• All four arms: 48 PASS / 2 FAIL with identical result sets. The two failures are pre-existing hermetic-mode effect refusals, present in both arms.
• regen_stage0 --verify → regen_divergence_count=0, with the seed rebuilt from the changed sources (not a stale-seed self-verify).
• The corpus is the behavioral oracle: the required compile-clean gate and floor resolve every module through these two functions, and are green on this head.

Retained evidence

docs/plans/receipts/fill-composition-overlay-direction/ — nothing here asks a reader to trust a table cell:

• subject.tsv — commits, both binary sha256s, host envelope, run order, memory instrument, invocation, selection
• cohort.tsv — 50 rows: ordinal, entry, function
• base-r{1,2}.tsv / after-r{1,2}.tsv + full per-run stderr
• base-arm-revert.patch — the exact one-variable reversal; the base arm is this tree with that patch applied and nothing else
• derive_summary.py → summary.json — computed only from the four arm files; refuses on a missing arm; keeps the inclusive rewire_total_observation row out of the exclusive sum

Two things are recorded rather than smoothed: the import-string rewiring row's +12% mean is one outlier inside a 35% within-arm spread (its other after run sits inside the base range), and the discovery-path pair is reported as an unpaired observation because both arms were OOM-killed — --roster-from-discovery does not restrict discovery to the supplied --entry rows.

Test plan

• cargo test -p v1-compiler-tests symbol_index_fill_overlay — 2 passed
• RED control executed: merge direction flipped in the seed → both arms red on the winner assertion (left: "probe.underlay" right: "probe.closure") while still producing the right key set, so the witness measures the direction and not the union. Seed restored.
• regen_stage0 --verify → divergence 0, seed rebuilt from changed sources
• cargo fmt --all --check
• Exact-head CI green: build, regen, heal_generated_artifacts, floor

Note on where the unit control executes: nextest was removed from CI 2026-07-11, so v1-compiler-tests compiles in CI but does not run there. The direction control executes locally (output above); the in-CI evidence is the corpus green.

Scope

Fill-composition slice: delivered. Broader shared-entry-view program: open. The surviving assembly rows — import-string rewiring, the per-entry closure census, the per-root variant base — are different defects needing module-grain reuse, and are deliberately not in this PR. Neither is the Rust build-path work.

Worker attestation

• Title/body describe the current scope only
• Tests executed locally, both green and red arms, with output
• No secrets / surprise commits