fix(sse): stop PII sanitizer splicing OpenRouter metadata into content (#13488)

[diegosouzapw]

Closes #13488

Root cause

createSseTextTransform's sanitizeObject (src/lib/sseTextTransform.ts) classified any
string field not in a hard-coded METADATA_KEYS deny-list as "content" via
getFieldCategory's default branch. OpenRouter SSE chunks carry a top-level provider
string and delta.reasoning_details[].text/.format alongside delta.content on the same
choice index — since neither provider nor native_finish_reason nor
reasoning_details[].format were in the deny-list, all of them fell into the same
per-choice "content" sliding-window FIFO buffer in streamingPiiTransform.ts as the actual
answer text, producing interleaved/scrambled output on both sides. The two files carried
independent copies of METADATA_KEYS that had already drifted (one had
system_fingerprint, neither had provider).

Fix

• Consolidated the two drifted METADATA_KEYS lists into one shared Set, exported from
  src/lib/sseTextTransform.ts and imported by src/lib/streamingPiiTransform.ts; added
  provider, native_finish_reason, and format to it.
• Added a classifyField(key, parentKey) helper (replacing getFieldCategory at the
  internal call sites) that also disambiguates reasoning_details[].text — which is
  reasoning text, not answer content — from ordinary text/content/delta fields, by
  tracking the enclosing object's key through the recursive walk.
• Applied the same classifyField/METADATA_KEYS logic to streamingPiiTransform.ts's
  onFlush generic-fallback branch (clearDeltas/populateRemaining) so metadata fields
  aren't blanked to "" or repopulated with buffered answer text there either.
• Deliberately kept the "any string field not recognized as metadata defaults to content"
  fallback (rather than switching to a strict allow-list) — tests/unit/piiReproduction.test.ts
  ("THEORY-004: Data Loss in Unknown Stream Fallbacks") already encodes that non-standard/
  unrecognized JSON shapes must not silently lose their text, and a strict allow-list broke
  that existing, intentional behavior.

Regression test

tests/unit/issue-13488-pii-openrouter-metadata-splice.test.ts — 3 cases: the exact
OpenRouter repro from the issue (mode warn), the same shape under mode redact, and a
second concurrently-streamed metadata field (native_finish_reason) to close the family
rather than just the one field named in the report.

RED on unfixed code:

✖ issue #13488: OpenRouter top-level `provider` field must not share the content buffer
  AssertionError: content must reassemble byte-identical to input even with PII sanitization enabled;
  got "etheetherSaetheris the laretherlake in Finland."
  - 'LakeSaimaa is the largest lake in Finland.'

GREEN after the fix:

✔ issue #13488: OpenRouter top-level `provider` field must not share the content buffer
✔ issue #13488: mode=redact must not splice `provider` and `content` either
✔ issue #13488: a second metadata field (native_finish_reason) must not share buffers either
ℹ tests 3 | pass 3 | fail 0

Gates run

• npx eslint --suppressions-location config/quality/eslint-suppressions.json src/lib/sseTextTransform.ts src/lib/streamingPiiTransform.ts tests/unit/issue-13488-pii-openrouter-metadata-splice.test.ts → clean, exit 0.
• npm run typecheck:core → exit 0, no errors.
• node scripts/check/check-file-size.mjs → the only ✗ is open-sse/utils/stream.ts (pre-existing, untouched by this branch — confirmed unchanged vs origin/release/v3.8.51).
• node scripts/check/check-complexity.mjs → OK — 2824 violations (baseline 3218).
• node scripts/check/check-cognitive-complexity.mjs → OK — 1278 violations (baseline 1437).
• node scripts/check/check-test-discovery.mjs → OK — new test file discovered.
• Full existing suites for the touched area, run individually (DB-per-file isolation):
  tests/unit/sseTextTransform.test.ts (16/16), tests/unit/streamingPiiTransform.test.ts
  (28/28), tests/unit/streamingPiiInitialization.test.ts (1/1),
  tests/unit/pii-opt-in-default.test.ts (5/5), tests/unit/piiReproduction.test.ts (5/5),
  tests/unit/adversarialPii.test.ts (11/11), tests/unit/piiSanitizerIpv6.test.ts (18/18)
  — all green.

Existing tests aligned

None weakened or deleted. tests/unit/piiReproduction.test.ts's "THEORY-004" case (arbitrary
msg key in a non-standard JSON shape must still be treated as content, not silently
dropped) directly shaped the fix's design — kept the deny-list-plus-explicit-additions
approach instead of switching to a strict allow-list, specifically so that test's existing,
intentional contract keeps passing unmodified.