refactor(mitm): replace inspector utilities clean-room

[diegosouzapw]

Summary

• replace four previously derived Traffic Inspector utilities with independent clean-room implementations: conversation normalization, SSE parsing/merging, secret masking, and header sanitization
• preserve public exports and observable caller contracts across OpenAI Chat/Responses, Anthropic Messages, Gemini, wrappers, and WHATWG SSE framing
• document the four replacements as standards-based clean-room work while keeping the two separately classified passthrough surfaces unchanged
• add no THIRD_PARTY_NOTICES entry because no upstream notice or source was reproduced

Clean-room boundary

Work started from exact base 091589089cd134a94df9f6cdab9ba562b2cefd18 in an exclusive worktree. The previous bodies of the four target files were never opened/read and were replaced with minimal stubs before implementation. No git show, git diff, git blame, or git log was used on those bodies; no llm-interceptor repository/package or provenance report was read. Inputs were limited to local black-box tests, public callers/types, repo documentation, and primary WHATWG/OpenAI/Anthropic/Gemini/RFC 6750 documentation.

src/mitm/passthrough.ts and src/mitm/_internal/bypass.cjs are intentionally untouched; their Git blob hashes match the exact base.

TDD evidence

• black-box baseline before stubs: 44/44 PASS
• deliberate minimal stubs RED: 6/44 PASS, 38 FAIL
• final focused suite: 54/54 PASS
• focused downstream consumers: 48/48 PASS
• ConversationTab consumers: 14/14 PASS
• 1,000,049-character masking probe: integral Bearer *** redaction in one linear scan

Gates

• npm run typecheck:core: PASS
• npm run typecheck:noimplicit:core: BASE-RED; exact base reproduces the identical TS7006/TS702x set in unchanged open-sse/translator/response/openai-responses/pureHelpers.ts
• focused ESLint on all changed TS/tests: PASS
• full npm run lint: HOLD/infra; stopped after about 8 minutes under severe shared-host contention without emitting a diagnostic
• npm run check:docs-all: PASS (historical soft drift warnings only)
• npm run check:changelog-integrity: PASS
• npm run check:cycles: PASS
• Prettier and scoped diff whitespace check: PASS
• full test:unit / test:vitest: attempted but not conclusive under shared-host CPU/SQLite contention; runs were interrupted after unrelated timeout failures. Focused and direct consumer matrices above are green.

This PR remains draft pending normal CI and review.