Affected Repositories

• PyAutoPulse (primary, only)

Branch Survey

Warning (not a conflict): unregistered worktree .worktrees/pulse-checkin-2026-10-17 on checkin/2026-10-17. Pulse PR #68 (docs/receipts) is open and unmerged: branch from main, rebase after it merges.

Suggested branch: feature/likelihood-lifecycle-p0-pulse (worktree ~/Code/PyAutoLabs-wt/likelihood-lifecycle-p0-pulse/)

Implementation Steps

1. pulse/summary.py: SUPPORTED adds ("profiling-summary", 3) and ("walltime-validation", 1); validate() dispatches v3 to catalogue.validate(doc, version=3) and walltime to a new pulse/walltime.py; hardware()/group_label() learn host_class.
2. pulse/catalogue.py: _record gains a v3 branch validating likelihood {dataset_type, variant}, instrument, regime, setup {backend, device, operator, precision, transformer, threads} + setup_id, host_class, model_variant, measurement, class, status, superseded_by, day_zero, figures[] {kind, path, sha256}, decision_id, release_pin, measured_at; enumerations: ten dataset_types (imaging, interferometer, datacube, point_source_image, point_source_source, multi_galaxy, group, cluster, multi_dataset, weak), host_class ∈ {hpc_ral_cpu, hpc_a100, local_laptop, local_gpu}, measurement ∈ {runtime, breakdown, gradient, compile, memory_host, memory_device, parity, walltime_validation}, class ∈ {optimal, secondary, experimental}, status ∈ {current, superseded, archived}. validate(version=3): key = (likelihood, instrument, setup_id, host_class, model_variant, measurement); refuse a second current per key; refuse class: optimal without decision_id; refuse status: current on local_laptop/local_gpu; refuse any archived record in records[] (they belong in archive_record_count); superseded requires superseded_by resolving in-document; planned_cells[] must carry class.
3. pulse/ingest.py + check CLI: resolve every decision_id against decisions.yaml at ingest/check time (doc-only validator checks presence); receipt records schema and current/superseded/archived counts.
4. pulse/walltime.py (new): records keyed by (likelihood, instrument, setup_id, host_class, model_variant, sampler, seed) with wall_s, likelihood_evals, implied_ms_per_eval, reference_runtime_record_id, predicted_ms_per_eval, ratio_measured_over_predicted, cores, parallel_efficiency, release_pin, slurm_job_id, max_log_likelihood, parity_status ∈ {match, diverged, not_assessed}; all producer-declared; duplicate keys refused.
5. pulse/registry.py: supported_schema accepts walltime-validation@1; a registry fixture row exercises it (no live row here).
6. pulse/decisions.py + decisions/README.md: optional kind: designation with likelihood, regime, hardware_class ∈ {cpu_many_core, gpu_a100, laptop}, optimal, secondary[], evidence[], rule, supersedes; template appended; validation of the designation fields when kind is set.
7. pulse/setup_browser.js: for v3 snapshots the selector block (before today's Instrument selector, ~line 909) becomes Class (Optimal default | Secondary | Experimental | All) → Likelihood (dataset type → variant) → Instrument → Hardware class → Measurement; default filter class == optimal && status == current; History checkbox includes superseded; Coverage panel lists planned_cells with class: optimal lacking a current record; v2 eligibility/curated() heuristics (lines 150–198) bypassed for v3. v2 path unchanged.
8. pulse/board.py + dashboard.md: per-instance status lines show current/superseded/archived counts; walltime axis renders as its own table under Profiling Results.
9. Tests: fixtures tests/fixtures/setup_summary_v3.json, walltime_validation_v1.json, refusal fixtures (duplicate current, optimal without decision, laptop current, archived in records); new cases in test_catalogue.py, test_summary.py, test_registry.py, test_decisions.py, test_ingest.py, test_board.py, test_setup_browser.py (first <select> is Class with Optimal selected; superseded hidden until History).
10. Docs: REFERENCE.md lines 9, 36, 42, 99–101 corrected; new @3, walltime and designation sections; README mirrors; CHECKIN.md lookup sentence.

Trade-off: one versioned reader with a v3 branch, not a parallel module, keeps the v2 instance untouched and lets 0B migrate lens by changing one registry pin.

Key Files

• pulse/summary.py — SUPPORTED set and version dispatch
• pulse/catalogue.py — v2/v3 record and catalogue validation
• pulse/walltime.py — new walltime-validation@1 validator
• pulse/decisions.py, decisions/README.md, decisions.yaml — designation kind
• pulse/setup_browser.js, pulse/board.py — Class-first browser and board
• REFERENCE.md, README.md, CHECKIN.md — contract docs
• tests/ — fixtures and cases

Likelihood lifecycle phase 0A: profiling-summary@3 (class/status), walltime-validation@1, designation decisions and Class-first Pulse board

Type: feature
Target: PyAutoPulse
Repos:

• PyAutoPulse
  Themes:
• profiling
• dashboard
  Difficulty: large
  Autonomy: supervised
  Priority: high
  Status: formalised
  Consequence: judge
  Witness: pyauto-pulse check accepts a fixture profiling-summary@3 catalogue in which every record carries class and status, refuses a second current record on one (likelihood, instrument, setup_id, host_class, model_variant, measurement) key, refuses class: optimal without a resolvable decision_id, accepts a walltime-validation@1 document, and the rendered board's first <select> is Class with Optimal selected and superseded rows hidden until the History toggle
  Review-minutes: 30
  Unattended: never
  Epic: likelihood-lifecycle
  Phase: 0
  Parent: draft/research/autolens_profiling/likelihood_lifecycle_epic.md
  Filed: 2026-10-21
  Note: Pulse PR docs: link /prm close-outs and refresh receipts after 2026-10-17 merges #68 (docs/receipts) is open and unmerged on 2026-10-21; branch from main and rebase after it merges — not a gate.

Phase 0 of the likelihood-lifecycle epic (ledger: draft/research/autolens_profiling/likelihood_lifecycle_epic.md; bodies and architecture: likelihood_lifecycle_epic_report.md §3.1, §3.2, §3.6, §3.8, §4 phase 0; evidence: likelihood_lifecycle_epic_surveys/02_pulse_insight_contracts.md). Phase 0 is the spine and has five sub-issues, one per repo, issued in order: 0A PyAutoPulse (this prompt) → 0B autolens_profiling → 0C autolens_inference → 0D PyAutoInsight → 0E PyAutoBrain. Each later sub-issue is filed as its own prompt (likelihood_lifecycle_p0_<target>.md, Phase: 0) once 0A is shipped, because 0B's catalogue builder and 0C's exporter emit the contract this issue defines. Library first: Pulse is the contract owner, so it goes first.

Rulings assumed (report §5; the human confirms or amends at plan approval)

• R1 vocabulary and setup_id grammar as report §3.1 (<backend>_<device>_<operator>_<precision>[_<transformer>]; site/tier moves to host_class).
• R2 day-zero history cut is label-only; nothing is deleted or re-run in phase 0.
• R3 the provisional designations in report §3.3 are the phase-0 labels; phases 2–3 confirm.
• R7 host_class is provenance, never identity; local_laptop / local_gpu rows can never be current.
• R4/R5 enter as validator enumerations only (mixed precision allowed with class: experimental; compile is a measurement, not a class axis).

What (sub-issue 0A, PyAutoPulse only)

1. profiling-summary@3 contract (pulse/summary.py SUPPORTED, pulse/catalogue.py validate, REFERENCE.md): additive on v2 with a version bump. Record fields: likelihood {dataset_type, variant}, instrument, regime, setup {backend, device, operator, precision, transformer, threads} + setup_id, host_class, model_variant, measurement, class ∈ {optimal, secondary, experimental}, status ∈ {current, superseded, archived}, superseded_by, day_zero, figures[] {kind, path, sha256}, decision_id, release_pin, measured_at. Enumerations for the ten dataset_types, host_class, measurement (adds gradient, parity, walltime_validation) and class/status live in REFERENCE.md and the validator. Rules: exactly one current per (likelihood, instrument, setup_id, host_class, model_variant, measurement); class: optimal ⇒ decision_id present and resolvable in decisions.yaml; host_class ∈ {local_laptop, local_gpu} ⇒ status != current; archived records are excluded from records[] and counted in archive_record_count; planned_cells carry class so coverage is reported per class. v2 instances keep validating unchanged (autofit_profiling stays @2).
2. walltime-validation@1 evidence axis (new schema, producer autolens_inference): record keyed by (likelihood, instrument, setup_id, host_class, model_variant, sampler, seed) with wall_s, likelihood_evals, implied_ms_per_eval, reference_runtime_record_id, predicted_ms_per_eval, ratio_measured_over_predicted, cores, parallel_efficiency, release_pin, slurm_job_id, max_log_likelihood, parity_status ∈ {match, diverged, not_assessed}. All derived numbers producer-declared; Pulse validates and renders only. Registry accepts a second summary per instance (or a second instance kind) for this axis; REFERENCE.md states these are full-fit measurements.
3. designation decision kind (decisions/README.md template, decisions.yaml, pulse/decisions.py): fields likelihood, regime, hardware_class ∈ {cpu_many_core, gpu_a100, laptop}, optimal: <setup_id>, secondary: [<setup_id>…], evidence: [record ids], rule (one sentence in dataset-property terms), supersedes. The validator's optimal⇒decision_id rule resolves against these rows. No designation records are written in 0A (phase 3 writes them); the kind, template and test fixture are.
4. Board (pulse/setup_browser.js, pulse/board.py, dashboard.md): for @3 instances the selector order becomes Class (Optimal default | Secondary | Experimental | All) → Likelihood (dataset type → variant) → Instrument → Hardware class → Measurement; a History toggle reveals superseded rows; a Coverage panel lists Optimal cells that are not_measured; archived records are counted, never rendered. @2 instances keep today's browser. The check-in prose (CHECKIN.md) says research lookups read class: optimal + status: current unless asked otherwise.
5. Doc drift: REFERENCE.md :9, :36, :42, :99-101 still describe v1 as live; fix alongside the @3 section. README mirrors.
6. Tests: @3 fixture catalogue under tests/fixtures/ exercising every rule above (accept, duplicate-current refusal, optimal-without-decision refusal, laptop-current refusal, archived exclusion); walltime-validation@1 fixture; browser test asserting the first <select> is Class with Optimal default and that superseded rows are hidden until History.

Out of scope for 0A

The autolens_profiling catalogue builder and move-only PRs (0B), the autolens_inference exporter (0C), Insight's additive fields (0D), the Brain lifecycle template and conductor grid (0E), any designation decision record (phase 3), any re-measurement (phase 3) and the setup-baseline campaign closure (phase 3).

Boundaries (epic invariants)

Optimal is always a producer-declared class backed by a Pulse decision record; Pulse infers nothing, selects no fastest row and computes no ratio. Exactly one current per key, chosen by the producer under the day-zero rule, never by date. Sessions end at their deliverable; the human runs /prm.