feat: input-side token-saving upgrades + eval quality gate

[edubraqd]

Why

In agentic Claude Code the weekly token limit is dominated by input — the whole context (system prompt, tool schemas, prior prose, every tool result) is re-sent every turn; caveman's output compression is the smallest slice. These changes target the input side and add a quality gate so compression never silently trades away correctness. A token saved on the re-sent prefix is re-billed every later turn (compounds); a token saved on output is billed once.

What

1. Agentic-loop output discipline + gated reinforcement + one-shot nudge (24c8e03)

• skills/caveman/SKILL.md: new ## Agentic Loop section — result-first, no recap of files just read, plan once then deltas, one intent clause per tool batch. Cross-references Auto-Clarity so security/irreversible/multi-step stay full prose.
• caveman-mode-tracker.js: per-turn reinforcement now fires every 3rd turn (+ on mode (re)activation), not every turn — the counter lives in a separate flag file so the injected string stays byte-stable and keeps hitting the prompt cache.
• caveman-activate.js: statusline-setup nudge is one-shot (a .caveman-nudged marker) instead of re-appending ~90 tokens to the cached SessionStart prefix every session.

2. Eval quality gate — fidelity axis + two-gate accept rule (5681437, 5c7e19a)

• measure.py only counts tokens, which is gameable (a skill replying k scores -99% and "wins"). Adds evals/judge.py (LLM judge for rubric facts + deterministic verbatim check -> fidelity.json) and evals/gate.py (accept only if tokens hold AND fidelity holds; high-risk prompts like the git-rebase "rewrites history" warning allow zero fidelity loss).
• evals/prompts/rubrics.json for all 10 prompts; a committed fidelity.json baseline; tests/test_eval_gate.py (incl. the canonical "reply k -> REJECT"); .github/workflows/eval-gate.yml.

3. Opt-in PostToolUse tool-result trim (de7da10)

• caveman-trim-tool-result.js: when enabled, replaces oversized Read/Bash/Grep/Glob results (via the public updatedToolOutput) with a deterministic, lossless-first trimmed version before they enter context. OFF by default (CAVEMAN_TRIM_TOOL_RESULTS=1); wired via --with-trim, independent of the plugin (no double-fire).

4. cavecrew as the default locate route (e2e7076)

• Nudge the main thread to delegate locate-shaped work to cavecrew-investigator (its verbose Grep/Read stays out of main context); cap investigator/reviewer output at 25 rows.

5. Structural ultra + honest input/output split in /caveman-stats (b01efd3)

• skills/caveman/SKILL.md: the ultra level now collapses multi-sentence explanation into one causal chain and cuts transition sentences (not just word abbreviation), keeping the verbatim code-symbol/error-string exemption.
• caveman-stats.js: /caveman-stats reports the input-vs-output token split — output is the demo, input is where the weekly budget goes — and points at the input-side levers. Number formatting pinned to en-US so output is locale-deterministic.

Testing

• tests/test_hook_output.js (8), tests/test_trim_tool_result.js (15), tests/test_eval_gate.py (7), tests/test_caveman_stats.js (31) all pass.
• Trim hook adversarially reviewed (determinism, data corruption, fail-open) — no issues.
• Gate verified end-to-end on the committed snapshot: caveman saves ~50% tokens at 100% fidelity.

Notes

• Clean-room: hook mechanisms are from the public Claude Code hook docs.
• The Windows safeWriteFlag temp-leak fix is a separate PR (fix(hooks): prevent safeWriteFlag temp-file leak on Windows rename co… #578) — this branch deliberately leaves caveman-config.js untouched.
• fidelity.json was judged inline (an authenticated claude -p was unavailable in the author's environment); regenerate with evals/judge.py once CLI auth is set up.

🤖 Generated with Claude Code

[edubraqd]

Heads-up on item 3 (the opt-in PostToolUse trim hook)

Verified empirically on a live Claude Code session (logging wrapper on the hook): the trim hook currently has ~no real-world effect, for two reasons:

1. Claude Code does not apply a string updatedToolOutput when the tool_response is a structured object — and every built-in tool returns one: Bash {stdout, stderr, interrupted, …}, Read {type, file}, Grep {mode, content, numLines, …}, Glob likewise. The hook fires, extracts the text, trims it, and emits a valid updatedToolOutput string — but CC shows the original result unchanged. So it's a no-op for all four matched tools.
2. Claude Code already persists oversized results natively (~30KB → saved to disk + a ~2KB preview), more aggressively than this hook would.

The hook code is correct and unit-tested; the limitation is in the PostToolUse API surface, not the implementation. It ships OFF by default, so it harms nothing — but it won't save tokens on current Claude Code. A doc note was added (src/hooks/README.md). If a future CC honors updatedToolOutput for object results, the hook would need to return the object shape with a trimmed text field (not a string).

The other items in this PR (agentic-loop discipline, gated reinforcement, one-shot nudge, eval quality gate, cavecrew locate routing, structural ultra, /caveman-stats input/output split) are unaffected.