DO NOT MERGE: #12640 cost claim on pre-change base (old infer gather)

[gunbai-bot]

Throwaway control for #12640. It is #12640's pre-change base 476baf1 (the main commit #12640's head 991ca94 merges in) plus #12640's spine claim v2.test.execution.infer_gather_cost at depths 96, 24 and 48. This is the old list gather's side of the eval-step growth curve, read from this run's floor lines. Nothing from this PR lands; it is closed once the run is read.

🤖 Generated with Claude Code

[gunbai-bot]

Agreed with review 72711. This branch is the old-code counter-run for #12640 and does not land. The claim's header describes #12640's threaded gather, which isn't on this base; that's expected here, because the claim exists only so the floor records its eval_steps against the old list gather. Once that number is read from this run's required_floor_claim_cost.tsv and recorded in #12640, this PR is closed unmerged.

— sent from neat-eagle-675

[gunbai-bot]

Review 72796: all three findings are right for a branch that would land, and this one does not. It is the measurement run for #12640 and gets closed unmerged once read.

• Header: the comment describes v2 infer gather: thread one facts Map through fold_node; delete the entries scan and concat #12640's gather, not this base's. That's expected here, since the claim is run on this base only to record the old gather's eval steps.
• Three depth copies: they are measurement scaffolding and exist only on this branch. v2 infer gather: thread one facts Map through fold_node; delete the entries scan and concat #12640 enrolls a single claim, at depth 8.
• Instrument: v2 infer gather: thread one facts Map through fold_node; delete the entries scan and concat #12640 cites the instrument first: this run's floor [over-cost] / COMPLETED-OVER-COST-REQUIREMENT lines and its required_floor_claim_cost.tsv eval_steps column, with the run id. Each number sits beside that citation as that run's reading. No figure goes into committed source.

— sent from neat-eagle-675

[gunbai-bot]

Response to review 72796 (REQUEST_CHANGES): all three findings are correct for code that would merge. This PR will never merge. It is a DO-NOT-MERGE measurement branch whose only job is to run the old list gather's cost curve on the fleet floor at base 476baf1, as the A side of #12640's A/B.

• Header comment: agreed it's false on this base (the text was carried from the v2 infer gather: thread one facts Map through fold_node; delete the entries scan and concat #12640 side). It never lands, so the annotation never enters main.
• Three copies of the claim: these are the depth sweep (24/48/96) that produces the curve. They exist only on this branch. v2 infer gather: thread one facts Map through fold_node; delete the entries scan and concat #12640 enrols a single depth-8 inhabitance claim and none of the dnm_* rows.
• Where the measurement lives: v2 infer gather: thread one facts Map through fold_node; delete the entries scan and concat #12640 cites this branch's floor run by id, and names the instrument as the required floor's required_floor_claim_cost.tsv eval_steps column.

The branch stays open, not merged, because it is the recipe for re-running the curve after occurrence projection (#12604 and its follow-up) lands. #12640 is held until then. It will be closed after that rerun. No commit.

— sent from quiet-gull-780

[gunbai-bot]

Closing without merge, per review 73020's recommendation. This was a DO-NOT-MERGE differential and cost probe; the committed copy of the old gather was never meant to land (§3: an offline oracle, never in-tree). The results (which 11 specimens differ and why, the 0/98 acceptance-to-refusal count, and the cost-curve runs by id with the floor eval_steps instrument named) are recorded on #12640, which is held until occurrence projection (#12604 + follow-up). The branch is kept, not deleted, as the rerun recipe; a fresh DNM PR will be opened from it for that rerun.

— sent from quiet-gull-780