mind: record samplers-surface-autolens-tiers shipped; file the ranker bug

[Jammy2211]

Two pieces of bookkeeping that merging #174 / PyAutoBrain#216 did not do by itself.

1. Lifecycle — the shipped task had no record

The samplers lane-tiers task shipped (PyAutoBrain#216 as 31b810e, PyAutoMind#174 as 485d1161) while its prompt was still sitting in draft/, so the ledger read not started for finished work.

Records it to complete/2026/08/samplers-surface-autolens-tiers.md and removes the draft. Written with lifecycle.py record --apply, which also refreshed complete/index.md (962 records).

One wrinkle worth noting for the next person: the prompt was folded into the record by hand, not via --prompt. That flag resolves only against active/ (lifecycle.py:726), and this task went draft/ → shipped without ever being issued, so there was no active/ file to find. The record is byte-equivalent to what --prompt would have produced.

lifecycle.py check → OK. lifecycle.py index --check → OK.

The record deliberately keeps the findings that cost real time, so they don't get rediscovered:

• a mature-tier cell's identity comes from its run_search() declaration and not its path — six cluster/ leaves declare dataset_class="group";
• PyAutoBrain/tests/ sits inside the tenant firewall, so a new test file must avoid naming repos (this one takes tier labels from module constants);
• the two scope exclusions (_MULTI_START_*_BY_CELL, the autolens benchmark table) and why.

2. The ranker bug, now filed

draft/bug/pyautobrain/feature_ranker_ignores_header_keys.md — found while planning the samplers task, when the Feature Agent's top pick had to be rejected by hand against the registry.

The Intake Agent classified it feature/ at medium confidence. I re-homed it to bug/, because REFERENCE.md already states the contract it breaks:

The Intake (Conception) Agent writes these automatically … sourcing Difficulty: from the shared sizing faculty the Feature Agent also uses — so the value shown up front is the one the Feature Agent later acts on.

It doesn't. agents/conductors/feature/_feature.py mentions Difficulty once, at line 258, where it prints its own re-derived score. It reads no header key at all — which is how draft/feature/autonomy/10_scheduled_runs.md, declaring Status: blocked, Priority: low and an explicit Blocked-by: gate whose body says do not start, came back as the recommended next pick.

Sizing (small / supervised / normal) kept from the intake pass.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WSCA1h6cSKMNwFmWEsxuCL

---

Generated by Claude Code