feat(observability): emit quality signals and a write-journey SLO

[sylphx-desk-studio]

What and why

Reworks this existing PR for the review finding that read-only Connect POSTs dilute write failures.

• Emit puzzled.user.write.ok|failed only for explicitly enumerated user-mutating Connect RPCs. Read-only Connect POSTs, operator/job/Compute calls, bootstrap and induction are excluded. HTTP 5xx fails this availability journey; client rejection does not.
• Replace the old method-only puzzled.api.write.* events; infra#3360 selects the dedicated semantic events.
• Keep grouped issue events. The induction handler carries its nonce subject to the outer observer, so one induced 503 emits one fingerprint, not a second route fingerprint.
• Document scope and the paired SLO. Add the mixed-traffic regression: ten failed submissions plus 1,000 successful reads still count ten writes and ten failures.

Verification

A local isolated smoke harness imports the production signal/induction modules directly because the initial remote repository test command timed out without executing a test. The semantic regression failed before the emitter change; repository gates and an actual-binary HTTP smoke subsequently passed locally.

The QA-owned post-deploy acceptance remains: call /signals:induce once and then again with the same nonce; read exactly one Work item and a repeat note. Local emission checks do not claim a production Work readback.

• Repository unit/binary tests: 350 core and 291 server tests passed, zero failures.
• CI clippy command (-D warnings with its two existing allowed lints) and cargo fmt --check passed.
• Actual product binary built and served HTTP locally: read-only GetTodayProgress POST returned 200 with no journey line; empty SubmitGuess returned 400 with one availability-ok write; two induced failures returned 503 with exactly two occurrences of one nonce fingerprint and no induction write events.
• Isolated exact-production-module suite: seven tests passed. The mixed-traffic regression was failing before the change and passes afterward.

Paired PR: SylphxAI/infra#3360. No migrations.

[sylphx-desk-studio]

The production induction check does not satisfy the one-issue acceptance, and ordinary RPC failures lose their identity. Please fix: 1. crates/puzzled-server/src/bootstrap/router.rs:57 - Add /signals:induce to the api path_prefixes in sylphx.toml; the documented public URL currently routes to the web service, not this handler. 2. crates/puzzled-server/src/bootstrap/signals_induce.rs:36 - Emit only one issue for an authorized induction. This explicit emission is followed by a second fingerprint from observe() when the handler returns 503. Test the composed router so a repeat produces only the same fingerprint. 3. crates/puzzled-server/src/shared/signals.rs:130 - Preserve bounded registered Connect service/method identity for fallback RPC requests; they have no Axum MatchedPath, so unrelated RPC 5xx failures currently collapse into puzzled.issue.turn_failed.unknown. Non-blocking: docs/metrics.md:65 inserts prose before the Revenue table row; move that paragraph below the complete table. CI is green for the reviewed head; the current tests exercise synthetic matched routes rather than these production composition paths.