Groundedness gate: streaming path, generator level, and QA metadata still read the pre-#143 signal

[NathanTarbert]

The groundedness gate is enforced on some paths and signals but not others. Three consumers still read a signal whose meaning changed in #143.

1. generateStreamingResponse gates on suppress only

packages/outpost/ai/src/pipeline.ts (streaming path) applies the suppression swap and nothing else — no groundedness penalty, no charged-claim clamp, no calibration, no formatter, no disclaimer, and no escalation.

So identical text takes two different routes:

	non-streaming	streaming
"Bug confirmed. Root cause is X." (no invented identifiers)	penalty 0.35, clamped below ESCALATE, disclaimer added, human paged	streams unmarked, no disclaimer, nobody paged

This was covered incidentally while suppression was claim-driven. #143 moved claims to penalty-only and the streaming path kept checking suppress.

Not currently reachable in production — generateStreamingResponse has no production caller (apps/web/src/app/api/qa/route.ts uses the non-streaming path and chunks the finished text itself). It is a trap for the next caller, not a live bug.

2. The generator's confidenceLevel omits the claim clamp

packages/outpost/ai/src/generator.ts → classifyGroundedConfidence clamps to SUPPRESSED_CONFIDENCE_CAP when groundedness.suppress is set, but does not clamp on groundedness.unverifiedClaims.length > 0. pipeline.ts does both. So for a charged-claim response the two levels disagree — the generator can report a level the pipeline would never report for the same text.

3. The QA route's metadata event omits suppressed

apps/web/src/app/api/qa/route.ts streams the safe published text (correct — it inherits the gate), but its metadata SSE event carries confidence, sources and latencyMs only. The UI cannot distinguish "here is your answer" from "we withheld an ungrounded draft and escalated", so both render identically.

Root cause, and the fix worth making

These are three instances of one process failure across #143: a signal's meaning changed and not every reader was swept. Same shape as the autoSend bug fixed earlier in the same PR (penalty moved to the pipeline, autoSend kept reading the pre-penalty score).

Patching the three sites invites a fourth. The structural fix is a single parameterized test asserting the gate's contract holds on every path and consumer — non-streaming pipeline, streaming pipeline, queue handler, QA route — so adding a path or changing what a signal means fails loudly instead of silently. Additive, and it would have caught all three of these plus autoSend.

Found by the CR loop on #143 (rounds 2-3), each reproduced against the merged code.

[NathanTarbert]

Two more reviewers landed after this was filed. Both mutation-proved that several of the tests covering the gate cannot fail — worth folding into the parameterized-coverage work above, since these are the tests that would otherwise be trusted to catch a regression:

• pipeline-groundedness.test.ts — toContain(AI_DISCLAIMER) + not.toContain(AI_DISCLAIMER_ESCALATED) cannot distinguish AI_DISCLAIMER_REVIEWED, so flipping the neighbouring ternary branch still passes.
• pipeline-groundedness.test.ts — "reports the penalised level for an ungrounded but publishable response" cannot fail: 0.975 and 0.825 both classify HIGH, so the penalised classification is unasserted.
• pipeline-groundedness.test.ts — not.toBeCloseTo(0.675) also passes for a triple, zero, or NaN deduction; it is strictly weaker than the equality assertion beside it.
• qa-api.test.ts — toContain('CopilotKit') is vacuous: the route emits 8-char chunks, so the word is split and only the metadata searchResults payload satisfies the assertion. Proved by mutation with all 9 tests still green.
• queue/src/__tests__/ai-response.test.ts — the shadow-mode suppressed case never asserts the ESCALATION job is enqueued, and in shadow mode escalation is the only path to a human.
• ai-response.test.ts — beforeEachTest captures at collection time, not per test, despite the name.
• groundedness.test.ts — the documented substring-grounding leniency (copilotKitTextarea counted as grounded by a source mentioning copilotKitTextareaWrapper) has no corpus row in either direction. That is the only known false-negative in the sole suppressing signal, so it is the one most worth pinning.
• groundedness.ts — the penalty rates (0.35 / 0.15 / 0.03) are module-private and hard-coded in ~12 test sites, while the ceiling and threshold are imported by name. Exporting them would stop the tests drifting from the implementation.
• pipeline.test.ts — the Pathfinder fake omits searchAll, which the sibling file's fake provides.

Also flagged and worth its own issue rather than burying here: generateStreamingResponse has ~8 tests covering it and zero production callers, and none of those tests would fail if the missing gate steps were wired up — so the coverage gives no signal about the thing that is actually wrong with that path.