[Jammy2211]

Summary

PR 3b of #442 (likelihood-lifecycle phase 0B). This finishes what PR 3 (#448) deliberately left: nothing reads or writes the old results/ layout any more.

• Dashboard trends re-keyed on the v3 identity (<measurement>:<dataset_type>/<variant>:<instrument>:<setup_id>:<host_class>:<model_variant>). The old-id → new-id map is committed in catalogue/series_id_map.json. It is generated mechanically from catalogue/migration_day_zero.json by the new build_series_id_map.py, which has a --check. Of 151 pre-move ids, 107 map to v3 ids. The other 44 have no mirrored cell (archived host-less summaries, or unresolved rows) and keep their old id under legacy:. No history lost: 144 series and 158 points before, 145 and 159 after. Every old id appears under its new id with the same releases. The one extra series is a real A100 point-source breakdown that the old file-name pattern missed.
• README tables: the runtime table uses v3 cell names (e.g. imaging/rectangular/hst). The breakdown table keeps every row it showed before, 106 → 110: archived rows are labelled "(archived)", and the 4 new rows are 3 light_fixed rows plus the A100 row.
• Writers re-pointed: about 40 Python writer sites and 15 HPC submits now write into the mirrored tree, and 8 readers changed to match. Sweep rows and comparisons share one placement rule (scripts/misc/likelihood_runtime/sweep_layout.py): sweep.py passes the exact path as --output-path, and aggregate.py reads with the same function. All 308 sweep-matrix rows get distinct paths under their own instrument.
• results_layout.py deleted, along with every import of it.
• Guard test test_results_write_destinations.py fails if any script writes a results path outside the mirrored tree. It follows / joins, .joinpath, os.path.join, Path(...) arguments and roots aliased within the same module. A results/<dataset_type>/ path must name a real scripts/<dataset_type>/<variant>.
• dashboard/summary.json record ids change to the v3 ids; the mapping is in catalogue/series_id_map.json. The dashboard regeneration also rewrites timestamps across ~557 files, as every full re-render does.

Needs a real run (not verifiable without profiling; nothing was run)

• sweep.py default root: only --dry-run and --skip-existing exercised. --skip-existing no longer counts the stale unresolved rows left under the old runtime folder, so those cells will be re-measured.
• size_sweep_table, fixed_light_scaling_table and gradient_mode_library_ab find new rows only if the writer's payload facts match the reader's hints.
• The fixed_light witnesses (levers/l2/l3/s4/s4b) and the point-source *_summary_v* / image_plane_solved writers: checked only by computing the paths.
• The vmap probe writer and resolve_vmap_batch: consistent by construction, no probe run.

Scripts Changed

• scripts/misc/tooling/build_dashboard.py: v3 series ids, native mirrored-tree scan, pre-move sources attached through the map
• scripts/misc/tooling/build_readme.py: native runtime, breakdown, baseline and misc-group scanners; v3-keyed breakdown table that keeps archived rows, labelled
• scripts/misc/tooling/build_series_id_map.py (new): old → new series id map generator with --check
• scripts/misc/tooling/results_layout.py: deleted
• scripts/misc/likelihood_runtime/sweep_layout.py (new): one placement rule for sweep rows and comparisons; sweep.py and aggregate.py use it
• _profile_cli.py: --output-path, defensive --instrument, legacy landing and locate helpers
• 12 likelihood_runtime*.py leaves: vmap probe landing directory shared by writer and reader; interferometer and datacube leaves record their transformer; point-source summaries placed
• fixed_light witnesses, numba scaling/draws, memo_policy, nautilus_batch_capture, separation_rule, oracle_figure, bakeoff, memo_after_summary, misc/latent: re-pointed
• size_sweep_table, fixed_light_scaling_table, gradient_mode_library_ab, lens/solver/_corpus: readers re-pointed
• hpc/batch_{cpu,gpu}: 15 submits re-pointed
• Tests: test_build_dashboard (mirrored fixture, continuity); new test_build_readme, test_results_write_destinations, test_legacy_result_helpers, test_sweep_layout

Test Plan

• ruff 0.17.2 check and format --check
• every lint --check step, including build_catalogue --check --validate-with ../PyAutoPulse and build_series_id_map --check
• pytest scripts/misc/test/: 1548 passed, 6 skipped
• smoke: the 9 lint.yml scripts plus the 4 changed interferometer/datacube leaves
• browser setup-page test
• lychee (CI)

Validation checklist (--auto run — plan was not pre-approved)

• Effective level: supervised (header: supervised, cap: feature/large → supervised). Explicit human auto launch on 2026-10-24, so the ship checkpoint resolves to opening a PR. Judge tier: merge is a human /prm. Merge order: this PR, then PR 5, then PR 6.
• Plan: on the issue (feat: likelihood-lifecycle 0B — mirrored results tree + profiling-summary@3 catalogue #442 comment "Auto run … PR 3b, PR 5, PR 6"), written at start, unmodified since
• Gate: tests 1548 passed / 6 skipped · smoke 13/13 · review: an independent Opus reviewer first returned FINDINGS (1 blocking: sweep writer/reader mismatch for interferometer/datacube; 3 non-blocking: dropped README rows, guard gaps, stray delaunay_nn comparison). All four were fixed in b244831..94dc857, and the re-review returned CLEAN, with the witness basis-cited. One accepted residual: the guard doesn't follow a results root passed as a default argument. · Heart YELLOW within the launch acknowledgement: manifest drift (generated hooks 3, root AGENTS.md 1, shared-standards 1, workspace checkouts 1), PyAutoHands and PyAutoMind open PRs 7 days old, release validation incomplete. A mid-run RED (local PyAutoLens 2 commits behind) was cleared by a fast-forward and a fresh Heart collection, with the human's OK.
• Human: plan sound in hindsight?
• Human: diff matches plan (no scope creep)?
• Human: merge, amend, or reject — then log the outcome

Refs #442

🤖 Generated with Claude Code

https://claude.ai/code/session_01SeDHwsR2dnwCk8Btfhej2L