release: v1.2.0 — the audit remediation release

[ojassug]

Closes every finding in max_audit.md plus the architectural work that followed. 78 commits beyond v1.1.0.

Why 1.2.0 and not 1.1.1

Two reasons, both concrete:

1. The roadmap planned this as three releases — v1.1.1 "Green Tree & Correct Metadata", v1.1.2 "Data Loss & Corruption", v1.1.3 "Honest Instruments". This ships all three, plus the Scope Decision Gate answers (H2, M1, M11) and Phase 1c. Calling it v1.1.1 would label three releases' worth of work as the first of them.
2. It removes command-line surface, which a patch release is not permitted to do.

⚠ Breaking

• --risk-tolerance, --max-output-tokens, --max-latency-ms removed with their TOKENDAMPER_* variables — now a hard Unknown argument rather than accepted-and-ignored.
• MCP optimize_context no longer accepts riskTolerance (it gained targetReductionRatio, which does something).
• GatewaySessionStoreInterface requires getSession.
• The Gateway no longer reads TOKENDAMPER_MOCK_UPSTREAM or NODE_ENV.

Nothing functional was lost. No stage, validator or planner ever read those flags. A script passing them now fails instead of silently doing nothing — which is the point.

Highlights

	before	after
MCP optimize_context	guaranteed 0% no-op reporting success	69.1% at targetReductionRatio: 0.3
45-file Python bundle	0.00% (26 constraint failures across 14 items reverted all 45)	22.73%
src/core knapsack	unreachable	15 of 31 files pruned, 20,540 tokens saved
tokendamper bench	threw for every installed user	runs

Plus: structured provider content survives the Gateway (a tool_result block could ship as a bare string); a 0% result now reports whether a budget was in effect, whether any transform could reduce the language, and whether items were reverted; and the docs say what the validators actually check — bracket/quote integrity, not syntax validity.

Verification

• 606 tests, typecheck and lint clean.
• All six version consumers verified reporting 1.2.0 (src/version.ts is the single source) — checked, not assumed.
• npm pack --dry-run: tokendamper-1.2.0.tgz, 421 files.

One deliberate non-change

test/fixtures/bench/baseline.json stays at "version": "1.1.0". That field records when the baseline was measured (commit 0f7dbb9, 2026-08-09), not the product version — re-stamping it would claim a re-measurement that did not happen.

Known limitations, stated in the release notes rather than omitted

Elision reduces TypeScript/JavaScript and Python only; Gateway mode is experimental and saves nothing across turns by design; --target-reduction-ratio engages the planner but is not a proportional target; drift remains bundle-scoped.

🤖 Generated with Claude Code