test+docs(cms): internal pkg tests, mutation baseline, honest README

[jamestexas]

Summary

Round 5 of the audit-equivalent hardening — closes the largest remaining mutation-testing gap, documents the lived-mutant inventory, and updates the README to reflect the testing methodology honestly.

Stacked on PR #16 — when #16 merges, this PR auto-retargets to main.

Why this round, explicitly

After rounds 1–4 we hit 82.19% efficacy with three real bugs fixed. But there were three structural gaps the prior work couldn't address:

1. pkg/cms/internal was invisible to mutation testing. No test file linked it, so gremlins reported every mutant as NOT COVERED.
2. No public inventory of what the test suite can't kill. A maintainer or reviewer would have to re-run gremlins and interpret each surviving mutant from scratch.
3. No public framing of the testing methodology. README claimed "Well-tested" but didn't enumerate what that meant — and didn't make clear this is NOT an external audit.

What's added

1. Internal package tests (pkg/cms/internal/asn1util_test.go)

Direct table tests for `MarshalSequenceHeader`, `MarshalSetHeader`, `MarshalImplicitHeader` and `MarshalSafe` — same byte-exact boundary-table approach as `pkg/cms/length_boundary_test.go` from round 3, plus `TestConstants` to pin the documented size limits.

Metric	Before	After
`pkg/cms/internal` efficacy	(untested)	100.00%
`pkg/cms/internal` killed	0	18
`pkg/cms/internal` lived	0	0

Side effect: `pkg/cms` itself rises from 82.19% → 83.40% because gremlins now sees additional coverage flowing through the internal helpers when `pkg/cms` tests run.

2. MUTATION_BASELINE.md

An explicit inventory of the 44 lived + 23 not-covered mutants on `pkg/cms`, decomposed into four categories with one-line justifications each:

• A — Logically equivalent (cannot be killed without removing real code, ~15)
• B — Defensive bounds (effectively unreachable at scale, ~12)
• C — Error-message format strings (low security value, ~5)
• D — Genuine boundary cases (addressable, low return, ~10)

Plus promotion criteria for future PRs and an explicit "not independently audited" disclaimer at the top.

3. README "Testing & Hardening" section

Honest framing throughout — leads with "not independently audited." Tabulates every test layer with the file that implements it. Lists the four real bugs found and fixed by the methodology with PR links. Documents what the methodology does not cover (cryptographic primitives → stdlib, side-channels → not tested, HSM → out of scope).

Also fixes a stale claim in Limits — "SHA-256 only" → SHA-256/384/512 per RFC 8419.

What this PR does not do

• Does not raise the CI mutation gate (still 80%). The 80% floor is for regression detection, not aspirational targets. Raising it now would discourage refactors that temporarily dip efficacy by 1-2%.
• Does not claim 100% efficacy on `pkg/cms`. See `MUTATION_BASELINE.md` for why this is intentional — half the remaining mutants are unkillable without making the code worse.

Validation

• Full suite passes under `-race` (make test)
• `make mutation-test` reports 83.40% on `pkg/cms`, 100% on `pkg/cms/internal`
• `make govulncheck` clean

Test plan

• CI green across all matrix entries (after PR test+ci(cms): RFC-traceable tests, named threat suite, mutation gate #16 merges and auto-retargets this PR)
• New mutation gate continues to pass at the 80% threshold
• `MUTATION_BASELINE.md` rendered correctly on GitHub
• README "Testing & Hardening" table links resolve to actual files

🤖 Generated with Claude Code