release: v0.17.1 — the evals run for the first time, and the tails go to zero

[sshlg]

Wave-3 of the 2026-08-29 family audit — the AST tails to zero, one patch release.

What's in

• AST-05 — test/evals/RESULTS.md gains its first two dated rows (2026-08-31, haiku + sonnet): 24 fresh blind trigger probes against the family's 28 skill descriptions and 6 scenario runs scored line by line. Both models 11/12 on triggers; scenario lines 10/12 (haiku) / 9/12 (sonnet); every miss named, Method section states the protocol and its limits.
• AST-B — the Anthropic generator-evaluator citation (read 2026-08-30) lands in agent-evals §5a beside the doctrine it converged with.
• AST-A1 — nine-subsystem coverage map closed on the board: the ninth subsystem is Automation; all nine map to existing doctrine with a file:line each; the suspected identity/approval-policy gap is a split verdict — approval policy covered (governance.md), auth mechanics a named delegation (layers.md:78-80).
• AST-07 — README's drifted "twenty references" aggregate dropped (24 actual; per-skill counts stay).
• AST-08 — license: MIT in all four skill front matters, each re-checked with yaml.safe_load.
• AST-09 — $schema in both manifests, using the two schemastore addresses that return 200.
• AST-10 — the three memory references join the orchestrator's index table (11 of 11).

Receipts

• npm test → OK: agent-stack structurally valid (13 checks = 9 named + 4 per-skill, 4 skill(s), v0.17.1), PASS: plant_guard — 9 cases, PASS: installer — 11 case(s), residue "left nothing" on every suite.
• python3 test/evals_validate.py → OK: 12 trigger cases and 3 scenarios validate; --self-test green.
• claude plugin validate . --strict and plugins/agent-stack --strict → both ✔ Validation passed.

🤖 Generated with Claude Code