Harden token savings advisory reports

[ictechgy]

Summary

• Add cache-score amortization advisory fields using explicit user-supplied cache write/read multipliers only.
• Extend tool-prune defer-report with gross/net deferred-schema char/4 proxy accounting and retrieval-required boundaries.
• Add a benchmark report measurement-baseline contract that documents captured fields, claim-eligible fields, proxy-only fields, and future run-identity gaps without changing CSV schema.
• Refresh README/plugin/kit docs and changelog while preserving no-provider-call/no-hosted-savings-claim boundaries.

Validation

• PYTHONDONTWRITEBYTECODE=1 python3 -m py_compile context-guard-kit/cache_score.py context-guard-kit/tool_schema_pruner.py context-guard-kit/benchmark_runner.py tests/test_context_guard_kit.py
• PYTHONDONTWRITEBYTECODE=1 python3 -m unittest tests.test_context_guard_kit.ClaudeTokenKitTests -k cache_score
• PYTHONDONTWRITEBYTECODE=1 python3 -m unittest tests.test_context_guard_kit.ClaudeTokenKitTests -k tool_prune
• PYTHONDONTWRITEBYTECODE=1 python3 -m unittest tests.test_context_guard_kit.BenchmarkRunnerTests -k benchmark_report
• python3 scripts/sync_plugin_copies.py --check
• git diff --check
• PYTHONDONTWRITEBYTECODE=1 python3 scripts/prepublish_check.py --skip-tests
• PYTHONDONTWRITEBYTECODE=1 python3 scripts/release_smoke.py --timeout 20
• PYTHONDONTWRITEBYTECODE=1 python3 scripts/prepublish_check.py — 691 tests OK

Claim boundary

This PR is advisory/local-only. It does not add provider calls, bundled pricing defaults, native provider tool-search configuration, lossy compression, or hosted API token/cost savings claims.

[ictechgy]

Quad review + validation evidence

Local validation before PR / after R2 fix:

• python3 scripts/sync_plugin_copies.py --check — OK
• git diff --check — OK
• PYTHONDONTWRITEBYTECODE=1 python3 -m py_compile context-guard-kit/cache_score.py context-guard-kit/tool_schema_pruner.py context-guard-kit/benchmark_runner.py tests/test_context_guard_kit.py — OK
• PYTHONDONTWRITEBYTECODE=1 python3 -m unittest tests.test_context_guard_kit.ClaudeTokenKitTests -k cache_score — 3 tests OK
• PYTHONDONTWRITEBYTECODE=1 python3 -m unittest tests.test_context_guard_kit.ClaudeTokenKitTests -k tool_prune — 14 tests OK
• PYTHONDONTWRITEBYTECODE=1 python3 -m unittest tests.test_context_guard_kit.BenchmarkRunnerTests -k benchmark_report — 13 tests OK
• PYTHONDONTWRITEBYTECODE=1 python3 scripts/prepublish_check.py --skip-tests — OK
• PYTHONDONTWRITEBYTECODE=1 python3 scripts/release_smoke.py --timeout 20 — OK
• PYTHONDONTWRITEBYTECODE=1 python3 scripts/prepublish_check.py — 691 tests OK

Quad review loop:

• Codex R1: REQUEST_CHANGES, MEDIUM cache amortization risk accounting issue.
• Forge R1: APPROVE with LOW note on same cache read-premium risk; accepted and fixed with the Codex MEDIUM.
• Agy R1: APPROVE.
• Claude R1: full-diff run produced no usable output; re-run on R2 fix diff.
• R2 fix: cache-score now compares expected cached vs uncached relative cost and returns no_read_discount/high for write-cheaper/read-more-expensive negative-savings cases; measurement baseline now lists primary_tokens_measured.
• Codex R2: APPROVE.
• Claude R2: APPROVE.
• Forge R2: APPROVE.
• Agy R2: APPROVE.

PR CI:

• test-and-prepublish (3.11) — pass
• test-and-prepublish (3.12) — pass
• test-and-prepublish (macos-latest, 3.12) — pass

No unresolved CRITICAL/HIGH blockers and no accepted unresolved MEDIUM blockers remain.