fix(claude): expose native Fast selectors with 1M context - #7748

Candidate 2c42324f97fd07859babedeff03efb724b778738, rebased onto dev at 8dd330a3f5165ea1685ad5282c5215a56dbc09f2. Local validation used macOS arm64 and repository-pinned Bun 1.4.4.

• PASS: 178 tests across nine focused files, including actual loopback HTTP discovery, readable/hashed IDs, both eligibility rules, input ceilings, parser/endpoint/cache behavior, and representative routed Fast behavior:

node_modules/.bin/bun run test \
  tests/codex-integration/fast-row-listing.test.ts \
  tests/claude-integration/claude-model-info.test.ts \
  tests/claude-integration/claude-long-context-eligibility.test.ts \
  tests/providers/fast-row-ingress.test.ts \
  tests/claude-integration/claude-models-discovery.test.ts \
  tests/claude-integration/claude-gateway-cache.test.ts \
  tests/claude-integration/claude-messages-endpoint.test.ts \
  tests/codex-integration/fast-row.test.ts \
  tests/providers/cursor/cursor-fast-listing.test.ts

• PASS: node_modules/.bin/bun run typecheck, node_modules/.bin/bun run structure:check, node_modules/.bin/bun run privacy:scan, node_modules/.bin/bun scripts/file-size-ratchet.ts, and git diff --check upstream/dev...HEAD.
• PASS: cd docs-site && ../node_modules/.bin/bun install --frozen-lockfile && ../node_modules/.bin/bun run build — 593 pages and 84,893 internal links checked.
• NOT GREEN: node_modules/.bin/bun run test:changed — 3,582 passed, 30 skipped, one failed across 175 files. loopback companion listener > binds 127.0.0.1 on the public port, and only loopback is credential-free fails at tests/server/loopback-listener-integration.test.ts:1082 with ECONNREFUSED on the host-selected non-loopback interface. It reproduces at the same assertion with the sole production change reverted to upstream/dev; all production files then match the base. The candidate was restored and its original fingerprint verified. Baseline reproduction command:

node_modules/.bin/bun run test tests/server/loopback-listener-integration.test.ts \
  -t 'binds 127.0.0.1 on the public port, and only loopback is credential-free'

• Before the fix, the new focused cases produced three missing-combination failures. Independent review used gpt-6-astra, xhigh, with no material findings. Its five file hashes and patch fingerprint remain identical after the clean rebase; the upstream effort-clamp change was inspected and the affected focused tests passed on the rebased candidate.

• FULL SUITE COMPLETED, NOT GREEN: node_modules/.bin/bun run test on the same head completed with exit 1: 52,690 passed, 170 skipped, three failed across 2,617 files (22 runner lanes, approximately 7m53s). All three failing cases also reproduce with the sole production hunk reverted to upstream/dev. This is baseline-reproducible failure evidence, not a passing full suite or permission to expand this native discovery fix.

• Two additional full-suite failures are in tests/config/serving-runtimes.test.ts: a real child that marks bind propagates a later nonzero exit (expected 42, received null), and a delegated child receives the parent's SIGTERM and its status survives (expected 0 or 143, received null). node_modules/.bin/bun run test tests/config/serving-runtimes.test.ts produced 34 pass / 2 fail both on the candidate and with all production files matching upstream/dev. The candidate bytes were restored and verified. No service/runtime repair is included.

• No actual Claude Code/Desktop UI selection, real upstream Fast inference, or large-context capacity test was performed. The local HTTP test proves discovery only. Shared installation and services remain unchanged.

• Historical external approval blocker on 2026-10-26 (subsequently resolved): Cross-platform CI and Catalog async contracts show “awaiting approval from a maintainer,” action_required, and zero jobs. The author account has upstream READ access (push, maintain, and admin are false), so it cannot approve these workflows. PR Labeler and PR hygiene are pending; the target/readiness workflow is queued. No CI success is claimed.

• Review status: zero submitted upstream reviews and zero unresolved review threads. CodeRabbit explicitly skipped the draft; its green status is not a completed review. No new material review finding is available to repair. The existing independent gpt-6-astra / xhigh review remains tied to identical file fingerprints.

• At the contributor checkpoint, the branch was on that day's dev tip (zero commits behind). The original contributor readiness checklist was left at 1/4; it was not retroactively marked complete. At that checkpoint, the PR remained draft. CI execution approval and maintainer code review are separate states.

Head: 2c42324f97fd07859babedeff03efb724b778738
Runtime: Bun 1.4.4; macOS arm64
Full parallel lane: 51,952 pass / 147 skip / 3 fail (2,596 files)
21 isolated runner lanes combined: 738 pass / 23 skip / 0 fail
Total: 52,690 pass / 170 skip / 3 fail (2,617 files); exit 1

serving-runtimes.test.ts, candidate isolated: 34 pass / 2 fail; exit 1
serving-runtimes.test.ts, upstream/dev production baseline: 34 pass / 2 fail; exit 1

Review readiness checklist

Historical automation template: this PR is already merged under the documented maintainer-integration exception. These boxes are not retroactive author attestations and do not describe an active draft.

• Required local validation passed; commands, results, and any full-suite exception are documented.

• I pushed my PR to a recent dev commit (at most 10 behind; a maintainer may still ask for the exact tip before merge).

• I resolved all correct Codex and CodeRabbit findings.

• My PR is ready for review.