feat(feedback): implicit signals — copy and continue (response-feedback PR-2)

[philmerrell]

docs/specs/response-feedback.md §11 PR-2, off develop after #1142 and #1145 merged.

Why

Thumbs cover a few percent of turns and skew negative. §10's implicit signals cover nearly every turn at no UI cost, and the row family has carried the signal discriminator since #1142 precisely so this could land without a backfill.

What

• Rows: same F# family, own key — SK F#{session}#{message}#{kind}, GSI_SK F#{message}#{kind} — so a signal never collides with the thumb on the same message. signal: "implicit", kind (closed enum), an ADDed count, updatedAt, TTL. No session rollup: the rows are the read model.
• Route POST /sessions/{id}/messages/{messageId}/signals with {kind}; 204 once accepted, same ownership check and RESPONSE_FEEDBACK_ENABLED kill switch as the thumbs.
• SPA: the existing Copy (on clipboard success) and Continue clicks in the message actions rail call MessageFeedbackService.recordSignal, fire-and-forget, deduped per message per kind per page load. Nothing in the UI waits on it or shows a failure.
• Admin profile: feedback.implicit = {copied, continued} as messages touched per kind (a message copied three times counts once), null when none; its own line under the Feedback tile.
• Never summed: every thumb reader — the messages-list merge, the profile's up/down and turn-class buckets, the rollups — filters implicit rows out. Tests pin it on both sides.
• FEEDBACK_ROW_PROJECTION gains kind and count; the content-policy walk covers them.

What this does not change

• Two of §10's four signals are deliberately absent. Edit-and-resend has no affordance in the SPA to hook. Abandonment is deferred: a session that goes quiet after a good answer looks identical to one that goes quiet after a bad one, so it needs its own design (or §10's offline "dissatisfaction in the next message" classifier) before it earns a row. Both recorded in the spec's §13.
• No new flag, no CDK, nothing in the inference path.
• Eval sampling (PR-4), the author surfaces (PR-5) and the report-dialog hatch remain unbuilt.

Tests

• Storage (moto): three rows for a thumb plus two kinds under distinct keys, count reaches 2 on a repeat, the thumb untouched, thumb readers and rollups blind to the implicit rows, unknown kind refused, other user not owned, preview sessions write nothing.
• Routes: 204 with the call forwarded; unknown kind and free text 422; kill switch 404.
• Profile: messages-touched per kind, unknown kinds ignored, null when none; the existing never-summed test still holds.
• SPA (worktree run against a per-package link of the main install with the two preview libraries stubbed): message-feedback.service spec (+2: once per message per kind and failure never surfaces; no server index / unavailable short-circuits), message-actions spec (+1: Continue records and still emits), session-profile.util spec (+1), anatomy page spec — 70 passed across those files. ng build (AOT) clean.
• Full backend suite on this tree: 9024 passed, 3 skipped.
• Not exercised live in the browser.

🤖 Generated with Claude Code