Summary

Phase 2 of the polish series (closes the local leg of #54): validate the VRAM/vmap story before the PreOptimizationTimes campaign burns A100 hours. All 9 campaign cells now run green end-to-end on CPU (durable record: results/notes/vram_validation_2026_07_08.md); the sweep flushed three real defects, all fixed here. The A100 probe leg is staged (submits + ingest workflow) for when RAL is back.

Scripts Changed

• vram/config.py — resolve_vmap_batch(): runtime cells prefer a fresh cell-matching vmap_probe_<model>[_sparse].json over the static table (a stale table can no longer OOM a job whose submit re-probed); VMAP_BATCH_SPARSE (w-tilde rows, unprobed cells fall back to the dense value); PROVENANCE (dense table dates to 2026-05-24 — recorded explicitly); non-GPU backends clamp resolved batches to 3 (the A100 table OOM-killed laptop-CPU runs at batch 64)
• vram/probe.py — probe JSONs record the backend that produced them; resolution ignores probes from a different backend (CPU memory analysis says nothing about A100)
• _profile_cli.py — check_pinned() / record_pinned_check(): pinned likelihood/evidence drift warns and records (pinned_expected + pinned_drift in the result JSON — the surface PyAutoHeart's filed vitals leg scans) instead of crashing a multi-hour job; boundary rule in results/notes/design_lock_in.md (profiling records and flags; autolens_workspace_test adjudicates library correctness)
• likelihood_runtime/<class>/<model>.py (all 9 cells) — wired to resolve_vmap_batch with the winning source logged; pins converted to record-and-flag (eager ≡ JIT ≡ vmap cross-consistency asserts stay hard); imaging/{mge,pixelization,delaunay} pins made per-instrument (the old module-level HST constants failed any non-HST run by construction); probe basenames normalised to vmap_probe_<model>[_sparse].json
• hpc/batch_gpu/submit_probe_{fast,delaunay}_a100 — probe-only sweeps re-validating every campaign vmap cell (dense + sparse imaging) against current source; ingest workflow in vram/README.md
• test/test_vram_config.py — 17 new tests (sparse fallback, probe override, mismatch/corrupt/backend handling); 25/25 total
• results/notes/vram_validation_2026_07_08.md — sweep verdict, timings, defects; dataset/imaging/euclid/ + simulators dashboard row (auto-simulated during the sweep)

Test Plan

• 25/25 unit tests; ruff check + format --check clean; build_readme.py --check idempotent
• Local CPU validation sweep: all 9 campaign cells green (incl. sparse variants); probe pipeline + probe-over-table resolution verified live; drift fields verified in real artifacts

Validation checklist (--auto run — plan approved in-session)

• Effective level: supervised (header: supervised, cap: maintenance → safe)
• Plan: on the issue (maint: vram-first validation sweep + batch-size provenance (polish phase 2) #54), approved interactively at start
• Gate: tests 25/25 (downstream n/a — standalone leaf) · smoke repo-gate pass + full live sweep (six-workspace smoke n/a — no downstream script surface) · review CLEAN (every changed path exercised by the sweep) · Heart YELLOW, human-acknowledged at ship sign-off (issue comment): reason set organism-wide, none touch this repo
• Human: plan sound in hindsight?
• Human: diff matches plan (no scope creep)?
• Human: merge, amend, or reject — then log the outcome

🤖 Generated with Claude Code