[Jammy2211]

Summary

Phase 2 of profiling-organ-birth (PyAutoPulse#1): the organ skeleton. PyAutoPulse reads, validates and links the profiling-summary@1 feed each registered <lib>_profiling project publishes, pins every ingest to one commit, keeps a receipt, and renders the cross-project board + cockpit feed. Modelled on the PyAutoEyes skeleton (PyAutoEyes#2) and the spec PyAutoBrain/docs/research/profiling_inference_organs.md. Human decisions 2026-10-02: one task, two PRs — merge this one first, then PyAutoBrain (witness row + boards: pulse).

What is in the box

• registry.yaml — lens row only; repo is a body-map identity, path/github resolve from PyAutoMind/repos.yaml (no second catalogue, per spec).
• pulse/registry.py, summary.py, ingest.py, board.py, cli.py; bin/pyauto-pulse check | board | census | fetch [--instance K].
• Ingest resolves the branch to one commit (GitHub API), reads summary_path at that SHA (raw URL), validates the exchange contract (schema/version, envelope keys, unique ids, finite numerics, dates, safe evidence paths, axis/unit consistency, hardware identity on comparison pairs), writes receipts/<instance>.json and a last-good snapshots/<instance>.json. A failed fetch renders cached with the original evidence time plus the new error; nothing is re-aged by a render.
• Board (dashboard.md/.html, state.json, badge.json): one row per instance; transport integrity, evidence freshness (valid_until null → "freshness policy unspecified" + age) and scientific qualification shown as three separate notions; detail per instance = the producer's own comparisons[] (refused pairs listed apart with the reason, ratio hidden) and limitations[]; drift items carry a /profiling triage <key> action (kind: prompt, safety: scientific_judgement). No recomputed ratio, no cross-project ranking, no Heart verdict. state.json is organ: pulse, validated by PyAutoBrain/board/_state.py.
• First live render committed: lens @ autolens_profiling d9f385f0, 159 records, 145 comparisons, status yellow (157/159 records unqualified by the producer — healthy transport never turns unknown quality green).
• Workflows: lint.yml (ruff, pytest, check --offline, lychee), pages_dashboard.yml (state validated against Brain before deploy), dashboard_refresh.yml (push main / PR dry-run / nightly / repository_dispatch: pulse-refresh, concurrency pulse-main-writers).
• AGENTS.md (boundaries + the spec's never-list; generated marker blocks untouched), REFERENCE.md (registry schema, the v1 contract as read, receipt format), README.md.

API Changes

New organ package pulse and CLI bin/pyauto-pulse. No PyAuto library API touched. Cockpit feed state.json (organ: pulse, repo: PyAutoPulse) is new; the legacy organ: profiling label in autolens_profiling/dashboard/state.json is left alone (phase 3).

Test Plan

• ruff check . && ruff format --check . clean
• pytest -q tests — 76 passed (hermetic: no_network autouse; fixtures valid_empty / unsupported_schema / duplicate_id / unknown_provenance / axis_mismatch / changed_hardware / second_producer / trimmed live feed; unavailable + stale-cached as scenarios)
• bin/pyauto-pulse check --offline → check: OK; online check → check: OK (lens d9f385f0)
• python3 PyAutoBrain/board/_state.py state.json → ok
• three workflow YAMLs parse
• Human: enable GitHub Pages on this repo (source: GitHub Actions) — the first Pages Dashboard run fails until then (PyAutoEyes#2 precedent)

Full API Changes (for automation & release notes)

Added

• pulse.registry.load/validate/get, pulse.summary.validate, pulse.ingest.ingest/resolve_commit/fetch_at/write_receipt, pulse.board.collect/render_*/write, pulse.cli.main
• bin/pyauto-pulse — check, board [--offline] [--from INSTANCE=PATH] [--mind PATH], census, fetch [--instance K]
• registry.yaml (schema 1), receipts/, snapshots/

Removed

• none

Migration

• none. Follow-ups filed separately: the pulse-refresh dispatch sender in autolens_profiling/pages_dashboard.yml; the Brain board card + cockpit identity transition (phase 3).

Design notes (deviations from the issue plan, all spec-conformant)

• --mind PATH is consulted before $PYAUTO_MIND (Eyes order); an explicit --mind without repos.yaml is an error.
• Comparison endpoints are resolved through the id grammar <comparison_key>@<version> (the live feed stores release versions, not record ids).
• A refused pair (axis mismatch / changed hardware / missing endpoint) is refused per comparison; the rest of the feed stays valid. host counts as changed only when both records name one.
• Receipts/snapshots are rewritten only when something other than fetched_at changes, so nightly refreshes commit nothing.
• --from local reads write no receipt and show the checkout basename only.
• lint.yml checks out Mind (_mind) and Brain (_brain) so check --offline resolves the body map and the state validator runs.

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code