Cut the eval suite from 19 to 17 cases, one capability each

[TheGreatAxios]

Net −549 lines (248 added, 797 removed), 19 → 17 cases. Replaces the earlier "add four cases" framing: the four rescued cases should displace redundancy, not stack on top of it.

The redundancy, confirmed by fixture

Four cases shared tests/fixtures/demo-comparison and tested one capability — implement or modify an endpoint in a toy API until tests pass — at 40 turns each:

Removed	Kept
complex-jwt	complex-stock-gate — richest of the four (multi-file mutable state)
complex-pagination
complex-idempotent-orders

Also removed complex-bugfix and its buggy-service fixture. simple-health (20t) and hidden-contract-inventory (40t) already bracket implement-against-tests at both ends of the difficulty range.

Of the four rescued cases, two earn their place

Kept — nothing else validates these:

• complex-fee-authority — authority resolution when docs and config actively disagree with the import graph
• freename-list-dir — discovering an off-wire tool via tool_search

Dropped as duplicates of already-baselined cases:

• complex-rename-spread duplicates complex-rename-user (rename across call sites)
• complex-pointer-chase duplicates misleading-symptom (trace indirection to the real target) — which is also currently the only case that discriminates between models, so it's the one to keep

Cost

Removes three 40-turn and one 30-turn case, adds a 35-turn and a 12-turn: −103 turns of budget per model per matrix run, concentrated in the slow tail.

Re-baseline required before merge lands meaning

Cases are auto-discovered, so every per-model total in the Capability eval tracking doc ("18/19", "19/19") is now against a different suite and stops being comparable. After merge: label the suite as the successor to v2 and record a fresh matrix. Fold with CL-6886, which called for exactly this fold plus the env-bait grader hole.

README case table updated in the same commit (guardrail 12). No test depends on the removed case ids — lib.test.ts references them only as synthetic sampleCase() fixtures.

Refs CL-6959, CL-6886.

[linear-code]

CL-6959

CL-6886

[TheGreatAxios]

Closing as superseded by #577, which replaces the suite entirely rather than trimming it.

The case against continuing here is in #577's own numbers: the last full matrix was 92/95 with all three failures on a single case, and suite v1's 14 cases were 14/14 on every model. A suite nothing fails cannot tell us whether CL-6940 helped, which is the reason we need one. Cutting 19 → 17 does not change that; #577 goes to four graded cases with real discriminating power.

Per the operator: #577 is the suite for the current and next releases' baseline.