Mark the Anthropic engine tier as presumed broken

[lemur47]

Close-phase staleness sweep, floor eeaf986. Doc-only, no behaviour change.

What the sweep found

Five surfaces advertise the Anthropic engine tier with no qualification:

Surface	Claim
docs/quickstart.md	--engine anthropic listed as an option to run
docs/library.md	engine-tier table
docs/architecture.md	engine-tier table
docs/STACK.md	status column read "Shipped"
CLAUDE.md	agent instruction file — read automatically into context

The tier's pinned model reached its published retirement date on 2026-06-15. A reader following any of these gets an opaque API error with nothing indicating it is a known fault rather than their own key or network. CLAUDE.md is the worst of the five: an agent reads it and suggests the broken path.

Marked unverified, not corrected to "broken"

The retirement date is documented; no live call has confirmed the failure from here — no API key in the environment and the host is outside the sandbox allowlist. Asserting "this is broken" would substitute one overreach for another, so the wording is "presumed non-functional, not verified against a live endpoint."

Restoring the tier is separate, tracked work. This PR only stops the docs advertising a capability that probably isn't there.

Also

docs/library.md now records that kind derivation lives in one exported function (derive_kind(), alongside classification_mean()), so someone implementing AnalysisEngine produces axes and calls it rather than carrying their own thresholds — the divergence PR #110 fixed.

Checked and deliberately left alone

• No stale references to the og/pillow extra, or .wrangler, anywhere in the durable surfaces — PR Drop the stranded og extra and restore github-actions to Dependabot #111 orphaned nothing.
• docs/library.md already stated the derivation canon correctly, including the mean ≥ 0.25 conjunct. It was docs/DESIGN.md — the canon document — that had drifted, and Hoist kind derivation out of the engines + cross-engine parity test #110 corrected it. No further change needed here.
• skills/si-quick-check/SKILL.md states the mean bands without the max-axis rule. Left as-is: it is a prompt for human and model judgement, not a restatement of the code path, and it does not contradict the canon.
• docs/architecture.md:74 describes derivation without stating thresholds — accurate, so not touched.

Test plan

• Pre-commit hooks pass, including the classification gate and gitleaks
• Doc-only diff — 5 files, 8 insertions, 5 deletions

🤖 Generated with Claude Code