v2.6.0 — M13: audit remediation & the learning loop

[danielPoloWork]

v2.6.0 — M13: audit remediation & the learning loop

This release completes milestone M13: a full-repo council audit turned into 14 focused fixes,
whose centrepiece is a closed, mechanical, human-gated learning loop. MINOR — every change is
additive; nothing removed or breaking.

The learning loop, closed and mechanical

• record_run.py (learning: record_run.py — mechanized run records with a failure channel and rubric scores #172) — one command writes a run record; overrides derive mechanically
  from the manifest's interview provenance, with durable failures and rubric channels.
• lesson_audit.py (learning: lesson_audit — regression-against-lesson detection, dead-lesson report, rubric trending #173) — a report-only watchdog: regressions against a recorded lesson,
  dead lessons, and low rubric trends — the trigger to promote an advisory lesson to a gate.
• Review-time lesson capture (git: PR-template "Lesson:" field + one-time ledger backfill from merged PR bodies #174) — an optional PR Lesson: field (a merged PR body is
  owner-approved by construction), lesson_sweep.py to harvest it, and the historical
  L-0003 / L-0004 backfill.
• run-records schema gate (lint: gate the learning loop — validate run records against the recorder schema #175) — learning/runs/** moves inside the enforcement
  perimeter; a malformed record now fails CI instead of silently poisoning the auto-tuner.

Render & interview hardening

• Interview provenance recorded in the manifest — asked vs defaulted vs imported (interview: record asked-vs-defaulted provenance in the manifest (interview state block) #169).
• A spec-substance floor in validate_manifest — no more hollow repositories (render: add a spec-substance floor to validate_manifest (deterministic path has no floor) #170).
• No silent it/d4np namespace fallback (render: silent "it/d4np" group_path fallback defeats the GROUP_PATH required-field guard #163); domain overlays are actually applied for
  game / web / mobile projects (workflow: domain_overlays are applied by no engine — a game/web/mobile project silently gets the base pipeline #165).

Factory internals

• render.py --check (validate-only) plus a gate-executability lint (workflow: gate "runs:" commands are never executed and have drifted — add render.py --check + a gate-executability lint #164).
• yamlmini extracted as a loud loader that rejects the interview: Phase 5 has no "do you already have a spec?" branch — it always co-authors, with no import-and-validate path #153 truncation constructs (tools: extract the YAML loader into yamlmini.py and loudly reject the #153 silent-truncation constructs #166);
  a leading UTF-8 BOM no longer breaks manifest parsing.
• The lint reporter is threaded through every check (lint: thread the failure reporter into checks — remove the module-global failures list #167); a test discovery runner means new
  tests need zero CI edits (tests: CI test enumeration is hand-maintained in two places and ungated — add a discovery runner #168).

Docs

• The Recall/Record steps are wired into the generation playbook (generate: wire Recall (Step 0.a) and Record (Step 9) into the playbook — the compliant agent never touches memory #171).
• README sweep (docs(readme): complete the tools table, add "How EADOS learns", date-stamp the model claims #176): a grouped tools table (all 27 tools), a new "How EADOS learns"
  section, and date-stamped model claims (the stale "Fable 5 not yet available" is fixed).

Full changelog: see CHANGELOG.md → [2.6.0].

---

This discussion was created from the release v2.6.0 — M13: audit remediation & the learning loop.