Measure repeated typechecking across selected entry closures

[gunbai-bot]

Summary

Slice-2 measurement-only instrument for entry-graph-union-construction (lane ci-cost). Resolves production-selected or explicit entries against one shared MultiEntryIndex and attributes typed-cache hit/miss/refused per (entry, typed module content key).

Decision metric: decision_ratio = repeated_typecheck_compute_ns / total_typecheck_compute_ns (cache-hit ratio diagnostic only).

Operator verdict (bounded N≤50, taken on candidate head): Close entry-graph-union-construction as a no-go for union construction on the shared-typecheck hypothesis. Repeated typecheck compute is zero at every measured cell (three disjoint 50-entry windows + 7-entry reorder fixture). Do not begin union construction. Redirect to per-entry assembly decomposition (new hypothesis, new denominator).

The candidate-head measurements have already taken the bounded operator verdict. After merge, one representative 50-entry cell plus the reorder control will be rerun on the merged SHA as a provenance/integration receipt — not as a second decision gate. Then delete one-shot measurement orchestration per receipt §G.

Instrument (dissolves post-merge)

• Flag-gated probe in reconcile_with_typed_cache / try_reconcile_all_cache_hits
• Bin: measure_repeated_typecheck_attribution — receipt JSON (--receipt-out); --entry-offset + --max-entries for disjoint windows
• Scaffold: cli_run_repeated_typecheck_attribution_probe — dissolution intrinsic to this lane (merged-SHA provenance receipt → delete); not Serve exact-tree resolved graphs from verified cross-process materialization #7534

Production receipts (N≤50, committed)

cell	offset	decision_ratio	repeated misses
narrow-disjoint	0	0	0
typical-disjoint	117	0	0
broad-disjoint	235	0	0

Explicit reorder (7-entry): decision_ratio=0, identical distinct-compute set both orders.

Full matrix + M2/retention context: docs/plans/entry-graph-union-slice2-typecheck-attribution.md §F–§G.

Structural law (retained post-deletion)

Within one shared typed-cache authority: one content key computes once; later requesters hit; entry order does not change the distinct-computation set. Enrolled in v1-compiler-tests (repeated_typecheck_attribution_*).

Controls (executed)

• Integration: shared-prefix hits, reorder invariance, partial all-hit-probe abandonment
• Arithmetic: decision_ratio aggregation semantics (deletes with instrument per §G)

Test plan

• cargo test -p v1-compiler-tests repeated_typecheck_attribution
• cargo test -p v1-compiler repeated_typecheck_attribution_arithmetic
• CI floor green on HEAD (8fc28f)

[gunbai-bot]

Addressed both findings from review 45863 in b1c912c:

1. Bin module docs (measure_repeated_typecheck_attribution.rs:10-15) — Fixed. decision_ratio is now documented as repeated_typecheck_compute_ns / total_typecheck_compute_ns; cache_hit_ratio is explicitly labeled diagnostic-only. Matches cli_run.rs (RepeatedTypecheckAttributionMeasurement::decision_ratio at the struct field) and plan §B.

2. HAND-RUST scaffold receipt (cli_run.rs:16237-16250) — Fixed. Added the sibling slice-1 pattern: DELETE WHEN list, rg -c cli_run_repeated_typecheck_attribution_probe src/v1/stage0/src/cli_run.rs receipt (returns 4 while scaffold stands, must return 0 at deletion), and the seed-Rust / gunbc.ci_release_bins enrollment note. Verified: rg -c returns 4 on current HEAD.

CI rerun on b1c912c is in flight.

— sent from valiant-lark-50

[gunbai-bot]

Addressed review 45883 in a56453d:

1. Membership denominator mismatch (cli_run.rs) — Fixed. When --entry or --max-entries changes the measured set, membership stats (sum_closure_memberships, union_modules, fanout_by_module) are now recomputed via closure_overlap_membership_for_entries for the actual selected_entries, not the full production roster. Production-only path still reuses slice-1 overlap unchanged.

2. Hand-Rust receipt (slice2 scaffold) — Fixed. Scaffold block now cites ROADMAP row entry-graph-union-construction (lane ci-cost, slice 2) matching slice-1’s pattern, alongside the existing rg -c cli_run_repeated_typecheck_attribution_probe deletion receipt. Receipt doc header updated with the same ROADMAP binding.

— sent from valiant-lark-50

[gunbai-bot]

Addressing review 46081 in 2b53bfb:

Finding 1 (scaffold lifecycle): Agree the old DELETE WHEN: slice-2 union verdict is taken trigger fired when §F landed. Per DESIGN §4b dissolution-on-climb, the discriminating controls stay enrolled — repeated_typecheck_attribution_* in v1-compiler-tests and repeated_typecheck_attribution_arithmetic are the executing evidence that decision_ratio=0 stays real. What dissolves is the one-shot measurement orchestration (measure_*, render helpers, measure_repeated_typecheck_attribution bin), not the probe hooks. Updated cli_run_repeated_typecheck_attribution_probe scaffold + receipt header with a concrete new deferral: delete measurement orchestration on #7534 merge or operator archive of slice-2 receipts (whichever first).

Finding 2 (run_matrix.sh): Agree — a hand-written bash orchestrator with no .dag carrier was the wrong shape for a durable authority. Deleted run_matrix.sh; disjoint-cell reproduction is now documented as three direct measure_repeated_typecheck_attribution invocations in §E/§F (no shell transport).

CI on the prior commit was a cold-cache timeout on cargo check -p v1-compiler-tests --tests (same step green in ~53s on 085f36ed); this push retriggers.

— sent from valiant-lark-50

[gunbai-bot]

Addressing review 46101: valid — the three repeated_typecheck_attribution_* integration tests arm a process-global probe without the serialization the sibling typecheck_compute_count oracles already use.

Fix: added with_repeated_typecheck_attribution_receipt (same Mutex<()> receipt-lock pattern as with_typecheck_compute_count_receipt) and wrapped all three tests. Production measure_repeated_typecheck_attribution is unchanged — only one caller, no lock needed.

— sent from valiant-lark-50

[gunbai-bot]

Addressing review 46165: valid — measurement_from_rows missed the entry_offset_applied: Option<usize> slot added to into_measurement at 711341bb. Fixed by passing None there (fixture tests don't exercise offset). Also initialized the probe-buffer fields (all_hit_probe_*, observed_entry_module_keys) that the struct gained during accounting repair — without those the lib-test crate wouldn't compile either.

— sent from valiant-lark-50

[gunbai-bot]

Review 46181 — fixed in d200591

The conflict was real: docs/plans/entry-graph-union-slice2-typecheck-attribution.md recorded the operator no-go while dag/gunbc/roadmap_authority.dag still had entry-graph-union-construction as active_large with boundary text saying attribution was "NEXT".

What changed (canonical authority):

• Updated the entry-graph-union-construction row: boundary now records CLOSED / no-go at N≤50, path points at the slice-2 receipt doc, red_control/handback reflect executed witnesses and delivered artifacts.
• Added RoadmapAcceptanceReceipt for entry-graph-union-construction (criteria_digest=4442f95272011c70, derived to match node_criteria_digest; verified green by witness_acceptance_receipts_are_valid_and_unique).
• Added entry_graph_union_construction_acceptance_note documenting the adjudication.

The row is now accepted in the authority graph (removed from active_roadmap_nodes); the receipt scaffold and roadmap agree on the same fact.

— sent from valiant-lark-50

[gunbai-bot]

Review 46190 — fixed

Valid: gunbc_ci_v1_compiler_tests_compile_gate_timeout_minutes was declared as bare Int = 10 while the adjacent step-timeout authorities (gunbc_ci_floor_step_timeout_minutes, gunbc_ci_regen_step_timeout_minutes) already use std.types.Duration.

Fix:

• gunbc_ci_v1_compiler_tests_compile_gate_timeout_minutes: Duration = 10 (same numeric value; no workflow drift)
• ci_build_job_backstop_is_step_sum_plus_prelude added to ci_workflow_witness_test.dag so gunbc_ci_build_job_backstop_timeout_minutes() stays pinned to the exact step-sum + prelude including the Duration-typed compile-gate term

ci_workflow_witness_holds green by execution.

— sent from valiant-lark-50

[gunbai-bot]

review 46319 — fixed in e05bb32

Verified: the attribution tests (repeated_typecheck_attribution_*) only acquired with_repeated_typecheck_attribution_receipt while counter tests used with_typecheck_compute_count_receipt; both touch process-wide TYPECHECK_COMPUTE_COUNT, so parallel cargo test could race.

Fix: replaced the two mutexes with one UNION_RESOLVE_PROCESS_RECEIPT_LOCK shared by both wrappers (counter + attribution probe state). cargo test -p v1-compiler-tests --lib union_resolve_receipts — 9/9 green locally.

— sent from valiant-lark-50