lint: guard that sidecar [outcomes] conditions can grade; inventory of current offenders (#2586, partial)

[maraxen]

Partial progress on debt #2586, which stays open. This PR adds a guard, fixes the four offenders that only needed structural fixes, and lists four criterion amendments for review. It changes no [outcomes] condition.

The guard. tests/lint/test_sidecar_conditions.py tokenises every [outcomes].condition in every *.bth.toml. It fails when a condition uses a bare identifier that is not a key in that sidecar's result_schema, because bathos can never grade such a condition. It accepts the expression forms bathos evaluates: SQL keywords, comparisons, arithmetic, IN, and scalar functions. Known offenders sit in tests/lint/sidecar_conditions_allowlist.toml, each with a reason. The test fails on a new offender and also on a stale allowlist entry. Positive and negative controls are included.

Structural fixes in this PR (fbb244e)

None of these change a criterion: no condition, decision, threshold or hypothesis text is touched. They only fix TOML structure and schema declarations.

Sidecar	Change	sha256 before → after
scripts/protonpotts/dump_protonpotts_oracles.bth.toml	Declares the four bools its conditions use. The script already emits them (dump_protonpotts_oracles.py:378-381).	20191493e726a4bb7b96e17c1b64dd26914b9ac2a63cd1462baba3d189164e13 → a0ef6c6135a77eb88ed5473119ddf4a523d9cba4546e21713f81c4d783954c1b
scripts/parity/laser_layers_tolerance.bth.toml	Adds a [result_schema] with band (str) and the keys the script emits (laser_layers_tolerance.py:106-117, band at L152).	4d3869684553398bd14b503658bd46bc2a693f14900f0528910f8c6cfadf6282 → fa49ae38298ad75cc44c9d077c28337f55fdbd17324026bc6443aed394b548fb
scripts/validate/poe_energy_sanity.bth.toml	Fixes a syntax error: the bare list under [tags] becomes tags = [...] in [metadata], with the same values. It still has no [outcomes], so it stays ungraded.	n/a
scripts/benchmarks/bench_temperature_array.py.bth.toml	Resolves the duplicate cells declaration into one [result_schema.cells] table with .item, keeping the same keys and types. It now parses, but its conditions are still prose (below).	n/a

Two cited passes rest on the first two files: e1967527 (oracle dump) and 2747efa8 (laser_layers tolerance). Both records carry the old sidecar sha above, so the edit is visible against them. Neither criterion changed. 2747efa8 graded without any [result_schema], which suggests bathos grades emitted keys whether or not they are declared. That is unverified, and a declaration-only change cannot alter it either way.

Amendments that need your review (not applied)

Each of these changes what a run must emit or what passes, so each is a pre-registration amendment. All four stay on the allowlist until you decide. No graded run of any of them is cited in the repo (searched docs and .praxia/; the bathos ledger was not queried).

1. bench_dedup_hetero.py.bth.toml. Conditions are prose per cell ("≥ 1.5 at K=16, N=32", "at all K", "any K < 1.0"). Proposal: the script also emits flat keys speedup_k16_n32 and min_speedup_vs_pytorch, and the conditions become speedup_k16_n32 >= 1.5, min_speedup_vs_pytorch >= 1.0 and min_speedup_vs_pytorch < 1.0. The thresholds map exactly, so only the form changes.
2. bench_temperature_array.py.bth.toml. speedup_vs_colabdesign_sequential >= M*0.7 at M=4 names a key the script does not emit; it emits speedup_vs_colabdesign per cell (bench_temperature_array.py:497). "speedup ≥ 1.0 at all M" does not say which baseline it compares against (colabdesign or pytorch, L497-498). Proposal: emit speedup_m4_vs_colabdesign and min_speedup_vs_<baseline>, keeping the 2.8 threshold at M=4. You choose the baseline.
3. sprint23_capability_analysis.py.bth.toml. The script emits no results at all; it writes only a Markdown report. Its schema keys (hardware_targets_covered, capabilities_covered) are never produced. Proposal: emit those two counts, with conditions hardware_targets_covered = 4 (pass) and >= 3 (marginal). The completeness thresholds are yours to confirm.
4. recapture/pottsmpnn_to_eqx.bth.toml. Prose ("h/J scale correct, k_neighbors preserved, round-trip …"), and no results emission. A run_sanity_check exists (~L353), but its result is never emitted. Proposal: emit h_scale_ok, k_neighbors_recovered and roundtrip_ok, with conditions as their AND. "Scale correct" needs a definition from you (e.g. a max relative error against the PyTorch h/J).

The PR also carries the UV_PYTHON=3.13 CI pin from #211. Without it, CI on main cannot install torch.

🤖 Generated with Claude Code

https://claude.ai/code/session_01SGHJhq24fYihaMivnXNhro