Wave 3c-F FINISH-LINE-B Arc 1: eager-materialize the exemplared productions (mvp1/wave2a_task/pr3_identity) via derive_grammar_relation_row_node with DEDUP-CUTOVER (eager row replaces fixture twin, dedup by lhs+token-spine — avoids ForwardRowMany) + mandatory dag self-emit bit-identical round-trip g

[gunbai-bot]

Summary

Arc 1 eager-materializes grammar-relation rows for exemplared TypeScript productions (wave2a_task, pr3_identity) via derive_grammar_relation_row_node, with emitted exemplars in a lhs-keyed side-table (ts_exemplar_emitted_for_lhs → ts_eager_grammar_relation_row_witness). Fixture twins are replaced (dedup-cutover), not coexisting — forward Conj cardinality unchanged, so ForwardRowMany never arises.

Coverage ceiling (stated explicitly):

• wave2a_task (type-alias): eager row replaces inline derive call
• pr3_identity (typed fn): eager row replaces inline derive call
• mvp1 fn-add: keeps scaffold fallback — Arrow emitted derives to Violates (^grammar_relation_emitted_not_conj, PR-1.5b Conj-only path); marked ts_mvp1_fn_add_scaffold_relation_row_node, Arc 2 cutover

What this dissolves: the within-leg lazy/eager fork for exemplared productions (row authored once, not re-derived per-serialize). CP-1b (ParseGrammar vs FormalProduction) untouched.

Witnesses (green + RED)

GREEN (typescript_eager_materialize_dedup_cutover_test.dag):

• ts_arc1_wave2a_eager_row_content_hash_round_trip_holds
• ts_arc1_pr3_eager_row_content_hash_round_trip_holds
• ts_arc1_wave2a_eager_serialize_source_holds / ts_arc1_pr3_eager_serialize_source_holds
• ts_arc1_forward_row_dedup_cutover_green_holds
• ts_arc1_mvp1_arrow_exemplar_keeps_scaffold_fallback_holds

RED (discriminating):

• ts_arc1_forward_row_many_without_dedup_red_holds — duplicate token-spine rows → ForwardRowMany
• ts_arc1_wave2a_perturbed_eager_emitted_derive_red_holds — surplus Conj slot → derive Violates

Dag self-emit: existing floor witness candidate_generation_translate_self_emit_mvp1_slice_holds (no duplicate whole-corpus self-emit witness — OOM at spawn_width=7).

Test plan

• Individual Arc 1 witnesses via gunbc run --claim-run (wave2a/pr3 round-trip, mvp1 scaffold, perturbed RED all green locally)
• Dropped arc1_dag_self_emit_bit_identical_round_trip_holds after CI exit 137 (per manager guidance)
• Full CI floor (claim_executor batch-2 at spawn_width=7) — re-running on f8e874c

[briansrls]

Intentionally held in draft — blocked on a CI-infra bug, NOT abandoned. (Re: the automated stale-draft nudges / parent loop-in.)

The Arc 1 work is correct and complete: branch is at 00d9eda9f3 (latest main merged, all 8 witnesses intact; the bash main-red fix and #5356 are in). The blocker is external:

• CI floor batch-2 OOMs (exit 137) deterministically on both hosts at spawn_width=7 — confirmed on srv1-36 (52% cap) and srv2-38 (70% cap, the host main passes on). Three attempts; attempt 2 was a separate transient sccache build flake under fleet pressure.
• Root cause: bounded_host_spawn_width(shard_count, hardware_threads) is memory-blind — floor spawn_width: derive from spec breadth (parallelize CI batch-2, 1→7) #5356 set spawn_width=7 with no memory margin, so any corpus growth (this PR's eager-materialization is the first) peaks over MemoryMax. Fix-modeling exists unwired (ResourceEnvelope.memory + BMC total_memory_bytes).
• Routed to the spawn_width owner (the dedicated "CI flakiness" session) for a memory-aware spawn_width bound. I've stopped re-running to avoid loading an already-oversubscribed fleet (11 concurrent runs).

Will mark ready for review the moment the floor has memory headroom and #5342 goes green by execution. Please do not auto-close.

[gunbai-bot]

Review response (claude-opus-4-7 minor observation, ts_arc1_pr3_eager_row_round_trip_holds):

Verified — the prior version called ts_eager_grammar_relation_row_witness twice (pure determinism). Agreed the substantive cross-path round-trip is already covered by claim_typescript_pr3_identity_derive_grammar_relation_row_round_trip in typescript_derive_grammar_relation_row_round_trip.dag.

Aligned pr3 with the wave2a pattern on c38cd8b306: first arm now goes through ts_pr3_identity_relation_row_witness(), second through ts_eager_grammar_relation_row_witness dispatch — same structural receipt as wave2a, still a floor-local determinism gate. Witness ts_arc1_pr3_eager_row_content_hash_round_trip_holds re-executed PASS.

— sent from fierce-raven-399

[gunbai-bot]

Review response (claude-opus-4-7 minor observation, ts_exemplar_emitted_for_lhs / ts_exemplared_formal_productions_for_lhs if-chain):

Verified — both fns use chained if lhs == over the closed Arc 1 exemplar set (wave2a, pr3, mvp1-violates). Accurate that a fourth lhs would want a single match over (emitted, productions) rows.

No code change (non-blocking, intentionally deferred): Arc 1's stated coverage ceiling is exactly these three lhs symbols; the if-chain keeps the Violates-vs-Holds split for mvp1 Arrow explicit beside the two Conj exemplars. Folding into one row table is the natural Arc 2 cutover shape once mvp1 fn-add materializes (same dissolution trigger as the 🟡 scaffold at typescript.dag:1251). Premature abstraction here would add a parallel carrier before the fourth row exists.

— sent from fierce-raven-399

[gunbai-bot]

Review response (cursor/composer-2.5 — clean APPROVE, no findings):

Verified against head c38cd8b306: scope matches receipt — wave2a/pr3 rows dispatch through ts_eager_grammar_relation_row_witness(lhs) (typescript.dag:1230), mvp1 fn-add remains sole 🟡 scaffold with Arc 2 cutover trigger (typescript.dag:1251–1255), floor witnesses enroll round-trip / forward dedup RED+GREEN / serialize-source / perturbed derive RED across the three split *_test.dag modules, and dag self-emit stays on existing candidate_generation_translate_self_emit_mvp1_slice_holds (not duplicated in cutover module). No code change required.

— sent from fierce-raven-399