Affected Repositories

• autolens_profiling (primary)
• PyAutoPulse (one-line registry re-pin after the builder merges)

Branch Survey

Claim guard worktree_check_conflict likelihood-lifecycle-p0-profiling autolens_profiling PyAutoPulse → exit 0.

Suggested branch: feature/likelihood-lifecycle-p0-profiling (PRs 1–6 as stacked/sequential branches off it, e.g. -map, -move, -builder, -misc, -coverage).

Decisions settled in the plan

• R3 provisional Optimal rows — Pulse REFERENCE.md: class: optimal requires a decision_id; none exists before phase 3. Emit class: secondary with "provisional_optimal: true" in limitations; phase 3 flips them when designations are written.
• Mirror test location — repo tests live under scripts/misc/test/; the mirror test (test_results_mirror_scripts.py: every results/<dataset_type>/<variant>/ has scripts/<dataset_type>/<variant>/ and vice versa, archive/ excluded) goes there; PR 5 re-homes it with its group.
• Move PR vs. path indexes — the move breaks path lists in results/provenance_grandfathered.txt, result destinations in catalogue/script_routes.json / wiki_bindings.json, and ALLOWED_NOTES_DIRS in check_results_layout.py. Regenerate these mechanically from the map inside the move PR (paths only, no measurement content; listed in the PR body) so main never goes red between PRs.

Implementation Steps

1. Map generator (script under scripts/misc/tooling/), emitting catalogue/migration_day_zero.json; retire legacy names in paths (pixelization → rectangular, *_numba → R1 setup_id, host/tier → host_class). Archive classes: dated suffixes, _lever, _probe, pre-2026-09 series, baselines/PreOptimizationTimes, frozen packs. sha256 recomputed after docs(cluster): record the human decision — keep the 1.0″ starting grid (#439) #441 merges.
2. Move PR applies the map with git mv only (plus map-derived path-index rewrites); check_results_layout.py --check green.
3. build_catalogue.py → v3 shape per PyAutoPulse REFERENCE.md "profiling-summary v3" (metric/unit/values; no setups[]/selections[]); §3.6 status rule (29 planned_baseline + 3 reference_candidate → current where the host rule holds; others on mirrored keys → superseded + superseded_by; archived → counted); R7 local_* never current; dashboard/catalogue.json regenerated; test_build_catalogue.py extended.
4. PyAutoPulse registry.yaml lens → profiling-summary@3; pyauto-pulse check accepts it; receipt refetched.
5. scripts/misc/ polish.
6. coverage_matrix.py + parity.json convention documented in results/README.md.

Witness

test_results_mirror_scripts.py passes; build_catalogue --check current, every record carries class/status, exactly one current per key; pyauto-pulse check accepts the re-pinned lens at @3 and its receipt refetches; the coverage matrix lists current/superseded/missing/stale cells per likelihood × class × measurement.

Key Files

• scripts/misc/tooling/build_catalogue.py — v2 emitter → v3
• scripts/misc/tooling/check_results_layout.py, results/provenance_grandfathered.txt — layout gate
• catalogue/registry.json, catalogue/script_routes.json, catalogue/wiki_bindings.json, _script_routes.py
• results/README.md — artefact policy + redirect table
• PyAutoPulse registry.yaml, REFERENCE.md

Likelihood lifecycle phase 0B: autolens_profiling mirrored tree, day-zero migration map, profiling-summary@3 catalogue builder, misc grouping, coverage matrix, parity convention

Type: feature
Target: autolens_profiling
Repos:

• autolens_profiling
• PyAutoPulse
  Themes:
• profiling
• dashboard
  Difficulty: large
  Autonomy: supervised
  Priority: high
  Status: formalised
  Consequence: judge
  Witness: tests/test_results_mirror_scripts.py passes (every results/<dataset_type>/<variant>/ has a scripts/<dataset_type>/<variant>/ and vice versa, archive/ excluded); build_catalogue --check is current and emits profiling-summary@3 with every record carrying class/status and exactly one current per key; pyauto-pulse check accepts the re-pinned lens instance at @3 and its receipt refetches; the coverage matrix tool lists current / superseded / missing / stale cells per likelihood × class × measurement
  Review-minutes: 40
  Unattended: never
  Epic: likelihood-lifecycle
  Phase: 0
  Parent: draft/research/autolens_profiling/likelihood_lifecycle_epic.md
  Filed: 2026-10-21

Phase 0 sub-issue 0B of the likelihood-lifecycle epic (ledger draft/research/autolens_profiling/likelihood_lifecycle_epic.md; report §3.1, §3.6, §3.7, §4 phase 0; survey likelihood_lifecycle_epic_surveys/01_profiling_inference_layout.md). 0A (PyAutoPulse#70, merged 2026-10-21) defined the contract this issue emits: v3 records carry likelihood, instrument, regime, per-record setup + setup_id, host_class, model_variant, measurement, metric/unit/values, class, status, superseded_by, day_zero, figures[], decision_id, release_pin, measured_at; there are no setups[]/selections[] — planned_cells (with class) replace selections; archive_record_count is required (PyAutoPulse REFERENCE.md "profiling-summary v3").

Rulings applied (confirmed 2026-10-21)

R1 setup_id = <backend>_<device>_<operator>_<precision>[_<transformer>], host/tier moves to host_class; R2 label-only history cut — nothing deleted, nothing re-run; R3 the provisional designations of report §3.3 are the phase-0 class labels (provisional, confirmed by phases 2–3; no decision_id exists yet, so every provisional Optimal row is emitted class: optimal only once a designation decision exists — until then class: secondary with a provisional_optimal: true note in limitations, or the builder refuses; decide in the plan); R7 local_* rows are never current.

What, as separate PRs in this order (data guard: moves and content never in one PR)

1. Commit nothing new first; migration map — catalogue/migration_day_zero.json: old path → new path (+ sha256) for every tracked file under results/ (1,456 files, 326 MB, results/lens/solver/corpus 165 MB) and results/notes/ ledgers → wiki/campaigns/<dataset_type>/<variant>/; dead experiments (dated suffixes, _lever, _probe, pre-2026-09 series, baselines/PreOptimizationTimes, frozen packs) → archive/YYYY/MM/<slug>/. Legacy model names retired in paths (pixelization → rectangular, *_numba → the setup_id). The map is reviewed by the human before any move.
2. Move-only PR (git mv only, reviewed as renames): the §3.7 layout results/<dataset_type>/<variant>/<instrument>/<setup_id>/<host_class>/{runtime,breakdown,gradient,fit,parity,provenance}.* with history/<release_pin>_<date>/ for superseded files; results/lens/<component>/, results/misc/<group>/; wiki/campaigns/<dataset_type>/<variant>/; archive/YYYY/MM/<slug>/ with a README pointer at each old parent; results/README.md redirect table. Never under a live RAL job on a RAL checkout of this repo.
3. Catalogue builder (scripts/misc/tooling/build_catalogue.py, catalogue/registry.json, _script_routes.py): reads the mirrored tree, emits profiling-summary@3, converts the config grammar, assigns class/status by the §3.6 rule (the 29 planned_baseline + 3 reference_candidate → current provisional where the host rule holds; other records on mirrored keys → superseded with superseded_by; archived → counted), planned_cells with class, figures[] where fit.png exists; build_catalogue --check; dashboard/catalogue.json regenerated; Pulse registry.yaml lens pin → profiling-summary@3 (one-line PyAutoPulse PR after the builder merges) and the Pulse receipt refetched.
4. scripts/misc/ grouping polish: misc/test/ split by group, misc/searches/ orphan removed, one-off scripts to archive/, a README per group.
5. Coverage matrix tool (extend the profiling conductor's campaign grid diff or add scripts/misc/tooling/coverage_matrix.py): likelihood × class × measurement → current / superseded / missing / stale, reading the v3 catalogue and the §3.3 provisional table.
6. Parity witness convention: parity.json per cell (reference log-likelihood, tolerance, status ∈ match / diverged / not_assessed) documented in results/README.md and read by the builder into measurement: parity records.

Out of scope

Re-measuring anything (phase 3); designation decisions (phase 3); fit.png rendering in _profile_cli.py (phase 1B); runtime_estimates.yaml (phase 5E); autolens_inference (0C); Brain conductor grid (0E).

Blocked-by (operational, not a header gate)

(Resolved 2026-10-22: that task merged as autolens_profiling#438 and the claim is released.) autolens_profiling was claimed by cluster-pointsolver-robust-settings (issue #436, branch feature/cluster-pointsolver-robust-settings, 5 unmerged commits, RAL sweeps live on 2026-10-21). The move-only PR cannot land under that branch or under live RAL jobs on a RAL checkout; start after that task ships (or the human releases the claim) — /start_dev registers this in planned.md until then.