[tae898]

Tooling for the post1 rerun of the paper's embedded cells (CAMPAIGN section 7 row 93 is the plan, #581 and PR #582 the docs row; I do not repeat it here). Draft: I am not merging anything, and nothing was run on the mini or in docker.

Closes #583

What it adds. A build-only mode (runner.py --build-only, one guarded hook after build() in each of l1_tpc.py, l2_graph.py, e2_hybrid.py, l3d_dense.py; the row carries row_kind = "build_only"), a wheel label on the rows of a stage that opts in (wheel_label, wheel_name, wheel_sha256), a per-cell choice of one wheel at landing (land_stage.py --post1-reps N), an embedded-only stage generator with a dry run, an estimate and a selftest (make_post1_rerun_stages.py), the procedure (POST1-RERUN.md) and a rollback by wheel label (post1_rows.py strip). A build-only row may replace only build_s, ingest_s, index_s, index_before_load, setup_s, gav_build_s and the load-transport facts; every timed figure of the cell stays at its 26.11.1 value, and the build window's memory figures sit beside the full row's as build_window_*.

Why the changed files are inert for the live pass. Comparator stages pull main at every start, so each edit has to be a no-op unless asked:

file	change	with nothing set
runner.py	import post1_rows, --build-only (default off), BENCH_BUILD_ONLY forwarded only when set, post1_rows.stamp_row(row) before the row is written	the import has no side effects, the flag does nothing, stamp_row returns the row untouched (no BENCH_WHEEL_LABEL, no BENCH_BUILD_ONLY)
l1_tpc.py, l2_graph.py, e2_hybrid.py, l3d_dense.py	if os.environ.get("BENCH_BUILD_ONLY") == "1": and a lazy import inside it	never true
land_stage.py	--post1-reps (default 0), and a refusal of rows that carry row_kind == "build_only"	no existing row has row_kind
engine_pin.py	jar_pin(), used in the two engine-jar comparisons	identity for every pin without .postN
CAMPAIGN.md	one paragraph before the paper-profile paragraph (no row, so it does not collide with row 93 in #582)	docs only

make_2611_stages.py is not changed (its --check still exits 0), so stage_drafts_2611.sha256 holds.

One finding. engine_pin.verify demands that the engine jar version equals the pin, and the pin of a post1 wheel is 26.11.1.post1 while its jar is 26.11.1: without jar_pin the runner refuses every post1 cell.

Dependency on #554. The native chain needs the four-arm generator of #554 (CHAIN_BUILD, the -native image tag); the generator refuses a native chain without it. I applied this branch onto the #554 head (git apply --3way): seven files apply cleanly, engine_pin.py conflicts in the two jar comparisons (same lines), and the same two-line change resolves it there. With that resolved, test_post1_rerun.py passes and a native and a JVM chain generate (44 scripts, all bash -n clean) on the #554 head too.

Checked. compile() of every changed file, pyflakes (no new warning), uv run --no-project python test_post1_rerun.py (resolve rules, the stamp, rollback, the generator: no served arm in any BACKENDS, --build-only and the wheel exports in every stage, the post1 rows file and rep state, no overlay in a build-only stage, a split cell list covers exactly the wanted cells, a drifted anchor is refused), bash -n on every generated script and on the driver, the runner's refusal of --build-only with the pass's own rows file (run as a plain argument error).

Not checked. No lane ran with the hook, no docker cell, no mini. The freeze's handling of the added columns and of two engine_version strings in one freeze (the one-version gate of the four-arm freeze) is for the landing's dry run. verify_pair_c25.sh may compare the wheel version with the jar version and was not read for .postN. The build_window_* copy matches field names by prefix (peak_, mem_, rss, cpu_, disk_, io_, heap, jvm_, gc_) and was not compared with real rows. The hours behind the dry-run estimate come from the scope note and the pass-1 build_s medians and are estimates, not measurements.

🤖 Generated with Claude Code