Score-contract tooling: audit_score_inputs, rule_probe --diff-lines, contract_pr_check, worktree_env, score-contract-audit skill (child of #2908) #2916

Tooling that #2908's efficiency assessment asks for before Phase 2, so that a score contract costs a fraction of a session the way a rule contract does (rule-contract-audit skill + contract-audit-scout). Each item is small and independent; a build-tier agent can do any of them from this text.

• tests/tools/audit_score_inputs.py <risk_metric> — generalise tests/tools/audit_documentation_inputs.py (risk_documentation score contract (draft) + per-file input decomposition audit (#2908 Phase 0) #2910). One adapter per _calc_* in an EQUATION_CASES-style table: which file_data columns feed which call arguments (the SHORT_KEY_MAP names: arch_api, def_doc, state_heat_triggers, …), how fid/irc/ot are obtained (_language_constants), how mp is obtained (_get_locational_multipliers(file_path)). Reproduce every recorded risk_<metric> from its recorded inputs; print one row per file, a CLUSTER table keyed by the dominant input, a SENSITIVITY table (score vs that input over its corpus range), and --summary (one paragraph: n files, residuals, dominant input, points-per-unit). Exit 1 on residual. Start with the five formulas that carry open-defect cells or a length leak: documentation, api_exposure, tech_debt, cognitive_load, verification.
• tests/tools/rule_probe.py --diff-lines before.json after.json [--context 2] — list lost and gained matched lines per corpus with counts, multi-line matches shown with context. The c api rule counts top-level variable declarations as public surface (a.c reads api 5 for 3 functions; +25.0% sits at the band edge) #2907 session wrote this twice; its first rule candidate lost 21 Doom K&R definitions that only this listing showed.
• tests/tools/contract_pr_check.py [--rules|--score] --corpus <kr-worktree> — run in parallel from the current worktree: the extraction tests for touched languages (git diff --name-only), rosetta_audit.py --allow-regressions, crucible_check.py, then bless_scope.py on the pre/post fixtures, audit_check.py; print a PR-body block with the four standard sections (Measured / Corpus / Golden masters / Tests) filled from the tools' own output.
• tests/tools/worktree_env.sh <name> [--corpus] — fetch; create or checkout --detach origin/main the engine worktree at ~/nyx_projects/gitgalaxy-worktrees/<name> (and the corpus twin); refuse inside the primary checkout; print the five exports (PYTHONPATH GITGALAXY_PATH GALAXYSCOPE_BIN LANGUAGE_CRUCIBLE_PATH KEYWORD_ROSETTA_PATH) as copy-paste lines. Every environment trap of the Epic: risk_documentation becomes a per-unit coverage ratio (Phase 4 score contract; child of #2812) #2908 session (wrong-path worktree, background job from a drifted cwd, stale corpus worktree producing a phantom mismatch) is this script not existing.
• .claude/agents/contract-audit-scout.md — the environment section names /srv/storage_16tb/projects/gitgalaxy/v6/.venv; on this machine the venv is ~/nyx_projects/gitgalaxy/.venv (full precision: .crucible_venvs/full_precision). Fix, and add the three tools above to its job list.
• .claude/skills/score-contract-audit/SKILL.md + docs/_score_contract_template.md — the score-layer twin of rule-contract-audit: same tier table and never-delegate list; phases = Epic: risk_documentation becomes a per-unit coverage ratio (Phase 4 score contract; child of #2812) #2908's with exact commands; the D-decision and acceptance-table templates; the contract-doc sections from docs/risk_documentation_contract.md.
• tests/tools/bless_scope.py --from-head — take the pre-bless fixture from git show HEAD: itself so the recipe is one command, and add --summary (files × leaf keys, one line each) for the PR body.

Child of #2908 (and of #2812 Phase 4, whose commensurability audit these feed).