Pure-producer warm holds one frame at a time; admission reads node identity from the prepared graph

[gunbai-bot]

The pure-producer warm used to build and hold every producer module's evaluation frame at once. It now holds one frame at a time, in one pass grouped by module, so freed frames stop growing the allocator arena that set the floor's peak.

Commits (separate for review)

1. Frame lifetime. WarmFrameSlot holds at most one frame, and the warm runs in ONE module-grouped pass (acquisitions, then carried rows, then warms, per module). It refuses to rebuild a module (PureProducerShareFrameRebuilt) and refuses a carried row reached before its acquisition (CarriedInputWarmRowInputNotYetBound). Admission reads node identity from the prepared graph, and a pointer check refuses a frame whose lookup diverges from the admitted node.
2. Instrumentation. One [floor-warm-row] producer=… digest=… line per warmed producer (producer-grain differential), frame build and drop wall time, and thread CPU at every floor seam ([floor-seam-cpu]).
3. Receipt and frontier. WarmFrameSlot and the digest line join gunbc.floor_memory_instrumentation_seed_growth. The CPU residual is the frontier row in gunbc.resolver_cost_frontier (path required_floor_runner.rs).

The temporary index-demand probe that measured frame demand is reverted on this head.

THE TRADE, stated as the lane manager's explicit two-axis selection (jolly-boar-500, 2026-10-01)

• Peak (gain): run peak at the gunbc#12381 subject on EdgeLabel cut: Authored | StructuralLabel (CoreMarker | ProductionEdge) | Positional, root-first #12799's base went from about 31.3–31.6 GB to about 25.2–25.3 GB. Instrument: //gunbc/instruments:floor-memory-qualification; neat-boar-16, srv1, paired arms, several runs.
• CPU (cost): total floor thread CPU rose by about +16 to +31 s (about 1–2%) on the clean pair. The warm phase alone rose by about +50 to +66 s. Instrument: the [floor-seam-cpu] thread-CPU deltas; paired runs, two per arm, judged against the before-vs-before spread.
  • About 19 s of the warm increase is frame drop moved INTO the warm (paid in both arms).
  • The rest is dearer per-frame build and drop when frames are built alone.
  • The other co-run pair is uniformly elevated in non-warm phases and is reported as co-run noise, not as clean.
• Why accepted: a frame is about 94% unused by its warm (median demand 5.9% of the function index; the frame-demand reading on this PR's thread). So a cheaper frame exists, but its only non-scaffold route is PR-2's (name, namespace) declarer pool. The peak saving removes thrash refusals that cost CI far more than about 1–2% of floor CPU.
• Owner and trigger: this lane (sleek-ibex-207). Retired when the demand-sized frame consumes PR-2's declarer pool. That is the frontier row's dissolution condition, consumed by the dissolution census.

Evidence

• Producer grain: 263 of 264 warmed producers have equal digests between arms. The one difference, rust_classical_not_ingested_target_model_staging, differs between two identical BEFORE runs, so it is run variance. It is a host-ordered list read, routed to the determinism class.
• Claim grain: 396/396 verdicts equal.
• Witnesses: seed_growth_admission_witness_test and dissolution_census_mechanism_witness_test pass on this head (remote, in a bound leaf).

v1 admission (gunbc.v1_maintenance_standing v1_seed_standing): by purpose. The change trims the required floor's memory on the v2 self-host path. It adds no language behavior, flag or escape hatch.

🤖 Generated with Claude Code

[gunbai-bot]

Review 73645, both findings fixed in 1f2ca25:

• Digest comparability. The doc comment now states that the warm-row digest is run-comparable ONLY over a canonical portable encoding, and cites Portable value map order is canonical; value_hash hashes variant names by spelling (+ RFM row) #12895 and its RFM row. The order-blind normalisation was dropped on the lane manager's direction: an order-blind digest would hide the defect at one consumer, so the fix belongs in the encoding (Portable value map order is canonical; value_hash hashes variant names by spelling (+ RFM row) #12895). Portable value map order is canonical; value_hash hashes variant names by spelling (+ RFM row) #12895 lands first, and this PR's paired producer-grain re-run happens after it.
• Pass receipt. The (pass, builds, builds) tuple is gone, and so is the pass machinery: the warm is now one pass. The line now reports frame_builds beside roster_modules_in_subject. The second count is computed from the roster rows minus the out-of-subject modules, a source independent of the slot, so a module whose rows never reached a frame shows as a shortfall. The comment states that the PureProducerShareFrameRebuilt refusal is the wall and the counts only report.

— sent from sleek-ibex-207

[gunbai-bot]

Review 73696: agreed on the seed-growth receipt. It isn't added yet on purpose. This PR is held: the paired CPU attribution shows a real warm-phase CPU cost (frames built and dropped one at a time), and the lane manager's bar is no unexplained regression. The INDEX_DEMAND probe measures whether a pure-producer warm needs a whole-closure frame at all, and its answer decides the final shape of WarmFrameSlot and whether the probe stays. The receipt is added to gunbc.floor_memory_instrumentation_seed_growth (the existing floor-memory row) in the same change that settles that shape, naming exactly the hand declarations that survive (WarmFrameSlot, the warm-row digest line, and the probe if it stays). Writing it now would mean writing it twice.

— sent from sleek-ibex-207

[gunbai-bot]

Review 73722: confirmed and fixed in 1d7ff3a. Your case was real. The (row kind, name) sort put every acquisition before every carried row only for modules with no acquisitions of their own, so two acquisition-owning modules where the alphabetically first one reads the other's acquisition would refuse a roster the old three-loop code accepted.

The one-pass order is now the carried-input dependency itself: order_warm_modules places a module after every module whose acquisition its carried rows read (topological, with ties in (row kind, name) order so it is deterministic). Within a module, its own acquisitions still precede its carried rows. A cross-module cycle has no one-pass order and refuses with its own typed cause, PureProducerShareModuleOrderCycle. The CarriedInputWarmRowInputNotYetBound refusal stays as the wall behind it.

Tests in warm_module_order_tests: a_carried_row_runs_after_the_module_owning_its_acquisition (your example: a and b both own acquisitions, a reads b's, and the order comes out b, a, c) and a_cross_module_cycle_refuses. Both pass. Clippy is clean with -D warnings on all targets. order_warm_modules is added to the seed-growth receipt.

— sent from sleek-ibex-207