samplers: scan the autolens findings-lane tiers

[Jammy2211]

Closes the gap the samplers faculty already admitted in its own AGENTS.md:

(Surface gap, filed: this faculty's SamplerSurface scan reads only the autofit-side tiers — it does not yet see the autolens experiment/mature tiers above.)

A conductor was told the findings maturation lane exists and then handed a digest that could not say where any search × likelihood combination sat in it.

Prompt: PyAutoMind:draft/feature/pyautobrain/samplers_surface_autolens_tiers.md

What changed

Two read-only, present-if-checked-out surfaces, matching the three that already exist:

• experiment (autolens_workspace_developer/searches_minimal) — the runnable probes, reusing _py_stems, plus the *_findings.md docs as name — first heading. The heading carries the campaign verdict, which is the reason a conductor consults the tier at all; a heading-less doc degrades to its bare name.
• mature (autolens_profiling) — the (sampler × dataset_class × model_type) cell matrix.

Resolution follows the existing convention (PYAUTO_LENS_DEVELOPER, PYAUTO_PROFILING); absent checkouts stay reported, never fatal.

The load-bearing detail: declaration, not path

A mature cell's identity is parsed from its leaf's run_search(...) call via ast, not from scripts/<dataset>/searches/<sampler>/<model_type>.py. The two genuinely disagree on the live tree: all six scripts/cluster/searches/*/mge.py leaves declare dataset_class="group" while their siblings declare "cluster".

group is legitimate, not a bug — each of those leaves passes an explicit default_instrument="hst", so the absent _DEFAULT_INSTRUMENTS["group"] key never fires. Path parsing would mislabel six cells and collide them with real ones. A leaf with no parsable declaration falls back to the path shape rather than vanishing.

Scope held

Inventory only, per the prompt — gaps stays keyed on the autofit promotion tiers, so no new judgment logic enters the faculty. Two deliberate exclusions:

• _MULTI_START_*_BY_CELL is not surfaced. The prompt names it as a way in, but it holds per-cell tuning knobs, not cell identity, and covers only 4 of the 34 cells.
• No autolens benchmark table. searches_minimal/output/ does carry a comparison.txt that benchmarks() would read verbatim, but d["benchmarks"] is single-valued and re-keying it is a schema break for no asked-for value. Worth a follow-up prompt.

Verification

Against the real trees: 37 probes, 8 findings, 34 cells, gaps unchanged, exit 4 (not a traceback) with no checkouts.

New tests/test_samplers_surface.py is hermetic — temp fixtures, no dependence on which siblings are cloned — and pins the declaration-beats-path rule with a fixture leaf whose declared class disagrees with its directory, so a future refactor cannot quietly regress to path parsing. It takes tier labels from module constants so the file names no repository and stays clear of the tenant firewall.

Two pre-existing failures in tests/test_skill_install.py are unrelated (confirmed on a stashed clean tree — they assert on installer output that differs in a cloud environment).

Companion

The tenant-firewall allowlist entries for these two files grow correspondingly in PyAutoLabs/PyAutoMind#174 — a new instance fact in an allowlisted file is drift, verified by running the check before the edit and seeing it flag exactly these two files.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WSCA1h6cSKMNwFmWEsxuCL