[Jammy2211]

Summary

Phase 0, sub-issue 0A of the likelihood-lifecycle epic (closes #69). Pulse gains the lifecycle vocabulary the four profiling/inference projects will share: profiling-summary@3 where every record declares its likelihood, setup/setup_id, host_class, model_variant, measurement, class (optimal / secondary / experimental) and status (current / superseded / archived); the walltime-validation@1 evidence axis for full-fit measurements from autolens_inference; the designation decision kind that an Optimal class must cite; and a Class-first, current-only setup browser with a History toggle and a Coverage panel. Pulse still validates and renders only: it infers no optimal setup, selects no fastest row and computes no ratio. v2 instances (autofit_profiling, the live lens pin) validate and render exactly as before; 0B migrates lens by changing one registry pin.

Decisions taken on the approved plan (flagged for /prm):

1. v3 records carry metric, unit, values instead of v2's measurement: {metric: value} object, because v3 uses measurement for the evidence-axis name and the one-current-per-key rule has no metric in its key (a breakdown keeps its components in one record).
2. v3 has no setups[] / selections[]: each record declares its own setup and planned_cells replaces selections. The plan called v3 "additive on v2"; it is additive in vocabulary and envelope, not in those two lists. Hazards/recommendations keep the v2 shape.
3. decision_id resolves against any decisions.yaml entry (no designation record exists until phase 3).
4. The v3 browser view is rendered in Python (selects and rows pre-built; JS only shows/hides) so pytest can assert the Class/Optimal default and the hidden superseded rows without a browser.
5. registry.validate now refuses any supported_schema the reader does not implement (live registry still valid).
6. A walltime instance raises no "unqualified" yellow item (that schema has no qualification flag).

Heart at ship: YELLOW — "manifest drift: workspace checkouts (manifest ↔ disk) — 1 mismatch(es) vs PyAutoMind/repos.yaml"; "release validation stale: source moved since rehearsal (PyAutoArray)". Neither concerns PyAutoPulse. Pulse PR #68 is still open; this branch is on main e3d411e with no drift, so no rebase was needed.

API Changes

pulse.summary.SUPPORTED adds profiling-summary@3 and walltime-validation@1; pulse.catalogue.validate(doc, version=3) enforces the lifecycle rules; new module pulse.walltime; pulse.decisions accepts kind: designation; pulse.registry.validate refuses unimplemented schemas; pulse.ingest and pyauto-pulse check resolve decision_ids and count current/superseded/archived; the board renders a v3 Class-first browser and a walltime table. No live registry row, snapshot or decision changed.
See full details below.

Test Plan

• ruff check / ruff format --check clean
• python -m pytest tests -q: 296 passed (207 on main)
• node tests/browser_setup.cjs and node tests/measurement_presentation.cjs pass (v3 page: first <select> is Class with Optimal selected; superseded row hidden until History)
• bin/pyauto-pulse check --offline OK on the live registry
• Witness by hand with a scratch registry: v3 fixture with a resolvable decision_id accepted (5 current, 1 superseded, 3 archived); unknown decision_id refused; duplicate current, optimal-without-decision, laptop-current and archived-in-records fixtures refused; walltime-validation@1 fixture accepted
• CI lint workflow green on this PR

Full API Changes (for automation & release notes)

Added

• pulse.summary.SUPPORTED — ("profiling-summary", 3), ("walltime-validation", 1); validate() dispatches both
• pulse.catalogue.validate(doc, version=3) — v3 record fields, enumerations (10 dataset_types, host_class, measurement, class, status, setup tuple + setup_id grammar), rules: one current per (likelihood, instrument, setup_id, host_class, model_variant, measurement); optimal ⇒ decision_id; local_laptop/local_gpu never current; archived refused in records[]; superseded_by resolves in-document; planned_cells[] carry class
• pulse.walltime (new) — walltime-validation@1 validator (keyed by likelihood, instrument, setup_id, host_class, model_variant, sampler, seed; producer-declared numbers; parity_status enum)
• pulse.decisions — optional kind: designation with likelihood, regime, hardware_class, optimal, secondary[], evidence[], rule, supersedes; template in decisions/README.md
• pulse.ingest / pulse.cli check — decision_id resolution against decisions.yaml; receipt counts current/superseded/archived
• pulse.board / pulse.setup_browser — v3 Class → Likelihood → Instrument → Hardware class → Measurement selectors, History toggle, Coverage panel, archived count; walltime table
• Fixtures: tests/fixtures/setup_summary_v3.json, walltime_validation_v1.json, four v3 refusal fixtures

Changed Behaviour

• pulse.registry.validate refuses a supported_schema the reader does not implement
• REFERENCE.md / README.md no longer describe v1 as live; new sections for v3, walltime-validation and designation decisions; CHECKIN.md: research lookups read class: optimal + status: current unless asked otherwise

Removed

• None

Migration

• v2 producers: no change. The autolens_profiling builder (0B) emits v3 per REFERENCE.md "profiling-summary v3" and the lens registry pin moves to profiling-summary@3; autolens_inference (0C) publishes walltime-validation@1 under its own registry row.

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code