Integration: all audit PRs merge clean and pass together (657 tests green)

[rldyourmnd]

All of the open audit PRs were combined into one integration branch and verified green together, so the set is safe to merge in any order.

Branch: rldyourmnd:audit/integration-all-prs (merges every PR below onto main).

PRs in the set:

• Docs / API: docs(readme): fix Library use example — package name and renderTextToImages #9, test(docs): guard README/docs relative links against rot #25, fix(vision): patch-based Anthropic image-token model + doc sync #27
• CLI / export: fix(cli): print the real package version for --version #11, fix(export): apply include/exclude and size limit to --git untracked files #13
• OpenAI: fix(openai): image array-form Responses system/developer content #15, fix(openai): record outgoingTextChars for Responses transforms #17, fix(openai): stop double-billing native GPT tool descriptions #19
• Applicability / hygiene: fix(applicability): match all proxy Anthropic message routes #21, chore(tsconfig): drop stale dashboard build comment and vestigial exclude #23
• Vision cost model: fix(vision): patch-based Anthropic image-token model + doc sync #27
• CI: chore(ci): least-privilege permissions and cancel-stale concurrency #29

Merge: all merged cleanly except one test-file overlap (tests/openai-gpt5.test.ts, where #15 added a Responses array-content test and #19 renamed the adjacent tool-schema test) — resolved by keeping both tests.

Verification (this integration branch):

pnpm install --frozen-lockfile
pnpm run typecheck   # clean
pnpm test            # 657 passed (33 files)
pnpm run build       # dist emitted; --version smoke check prints 0.8.0

Env: Node v24.16.0 / pnpm 10.21.0 locally; CI is Node 22 / pnpm (action-setup). One render-heavy e2e test (design-behavior-e2e) can time out at the default 5 s only under peak parallel load on a cold machine; it passes in isolation and on CI — raise --test-timeout if you see it flake locally.

Nothing to merge here — the individual PRs are the reviewable units. This is just the "do they compose" receipt.

[rldyourmnd]

Refresh (round 3): re-cut rldyourmnd:audit/integration-all-prs from main and merged the amended + new PRs. Everything composes green.

Added since the last refresh:

• fix(vision): patch-based Anthropic image-token model + doc sync #27 amended — exact Anthropic resize (binary-search reference), stale-comment fixes
• fix(prompting): discourage guessing exact strings from imaged text #32 — exact-string abstention wording in the rendered framing
• fix(factsheet): protect high-risk camelCase identifiers without evicting tier-0 #34 — camelCase factsheet protection (tier-1, no tier-0 eviction)
• test(eval): Opus 4.8 lower-density read sweep #36 — Opus 4.8 lower-density read-sweep harness (measurement only)

Full set: #9, #11, #13, #15, #17, #19, #21, #23, #25, #27, #29, #32, #34, #36 (14 branches).

Merge: clean except the same one test-file overlap in tests/openai-gpt5.test.ts (#15 vs #19), resolved by keeping both tests.

Verification:

pnpm install --frozen-lockfile
pnpm run typecheck   # clean
pnpm test            # 663 passed (34 files)
pnpm run build       # dist emitted; --version smoke check prints 0.8.0

Env: Node v24.16.0 / pnpm 10.21.0 locally; CI is Node 22. Individual PRs remain the reviewable units.

[teamchong]

sorry for the late follow-up, finally got all of these merged in. many thanks!