fix(security): random per-process self-loop admission bearer (#13679 PR C)

[diegosouzapw]

Refs #13679

Not covered here (the other 15 findings from the #13679 umbrella, shipped as separate PRs):

• item [WIP] Add FastAPI OpenAI-compatible LLM gateway with routing #1 cloud-sync HMAC fail-open (PR A)
• item build(deps): bump actions/github-script from 7 to 8 #2 CLI/management token salt (PR B)
• items feat(versioning): set initial version to 0.0.1 and restore dynamic model import #4/fix(oauth): prevent duplicate connections on re-authentication #5 deploy-manifest literal secrets (PR D)
• items fix(oauth): broaden upsert to match any existing connection #6/chore(env): sync .env.example with current .env structure #7/fix(oauth): prevent connection test from corrupting valid tokens #8 keyless-by-default + public listener posture (PR E)
• items feat(proxy): enable SOCKS5 proxy support by default #9/fix(sync): disable cloud sync and truncate error logs #10 Docker network isolation + debug listeners (PR F)

Root cause (short)

resolveSelfLoopBearer() in src/shared/middleware/chatAdmissionIdentity.ts fell back to the
checked-in literal "sk_omniroute" whenever OMNIROUTE_API_KEY/ROUTER_API_KEY were unset.
Anyone reading the source knew this shared secret and could send
x-omniroute-admission-bypass: internal + Authorization: Bearer sk_omniroute to skip the
heavyweight admission/queueing lease.

Traced every caller of isInternalAdmissionBypass(): it is consumed only by
admitChatRequest() in chatBodyAdmission.ts, and gates skipping the per-connection
admission/queueing lease — not authentication/authorization. A forged header pair only skips
capacity reservation (still subject to the hard byte-size cap); it does not bypass API-key/auth
checks elsewhere in the pipeline. So the real-world blast radius is narrow, but the predictable
literal was still bad practice worth fixing (lowest priority of the 6 PRs per the owner's
ordering).

Fix

Generate a random 32-byte secret (crypto.randomBytes(32).toString("hex")) at first use,
memoized in memory for the process lifetime, used only as the last-resort fallback when neither
env var is set. Both the in-process caller (audioBridgeHelpers.ts / visionBridgeHelpers.ts)
and the verifier (isInternalAdmissionBypass) call the same resolveSelfLoopBearer(), so they
still agree on the value within one process — no cross-process coordination is needed since this
bearer only gates a same-process self-loop call.

Regression test

tests/unit/chat-admission-selfloop-random-bearer-13679.test.ts

RED (on unfixed code):

✖ resolveSelfLoopBearer never falls back to the predictable sk_omniroute literal (15.552138ms)
  AssertionError [ERR_ASSERTION]: resolveSelfLoopBearer() fell back to the checked-in literal —
  a shared secret anyone reading the source knows, instead of a per-process random value
    actual: 'sk_omniroute'
    expected: 'sk_omniroute'
    operator: 'notStrictEqual'

✖ resolveSelfLoopBearer's generated fallback has enough entropy to resist guessing (1.903055ms)
  AssertionError [ERR_ASSERTION]: generated self-loop bearer is too short to be a random secret:
  "sk_omniroute" (12 chars)

GREEN (after fix):

✔ resolveSelfLoopBearer never falls back to the predictable sk_omniroute literal (11.636074ms)
✔ resolveSelfLoopBearer's generated fallback is stable within the same process (1.27507ms)
✔ resolveSelfLoopBearer's generated fallback has enough entropy to resist guessing (2.574413ms)
✔ resolveSelfLoopBearer still prefers OMNIROUTE_API_KEY over the generated fallback (0.878576ms)
ℹ tests 4
ℹ pass 4
ℹ fail 0

Gates run

• npx eslint --suppressions-location config/quality/eslint-suppressions.json <changed files> → clean, no warnings
• npm run typecheck:core → exit 0
• node scripts/check/check-file-size.mjs → 1 pre-existing violation on open-sse/utils/stream.ts (base drift, file not touched by this PR — confirmed on origin/release/v3.8.51 tip, 3114 lines vs frozen baseline 3098)
• node scripts/check/check-complexity.mjs → OK (2824 violations vs baseline 3218)
• node scripts/check/check-cognitive-complexity.mjs → OK (1276 violations vs baseline 1437)
• node scripts/check/check-test-discovery.mjs → OK (new test file discovered)
• DATA_DIR=$(mktemp -d) node --import tsx/esm --test --test-force-exit tests/unit/chat-admission-selfloop-random-bearer-13679.test.ts → 4/4 pass
• DATA_DIR=$(mktemp -d) node --import tsx/esm --test --test-force-exit tests/unit/chat-body-admission.test.ts → 37/37 pass
• DATA_DIR=$(mktemp -d) node --import tsx/esm --test --test-force-exit tests/unit/timing-safe-compare.test.ts → 6/6 pass (reads this file's source for pattern checks)

Existing tests aligned

tests/unit/chat-body-admission.test.ts had a test asserting resolveSelfLoopBearer() equals
the literal "sk_omniroute" when no env key is set — that encoded the old (insecure) contract.
Updated it to assert the fallback is NOT the literal and is memoized per-process, matching the
new corrected behavior. No assertion was weakened or removed; the other sk_omniroute-literal
tests (rejecting the sentinel once an env key is configured) were untouched since they test a
different, still-valid scenario.