fix(warehouse): label an exception-less error span from its span attributes

[Makisuo]

What

error_events_mv / error_events_by_time_mv derived the error label and fingerprint from the first OTel exception span event, else StatusMessage, else the literal Unknown Error. Cloudflare's native Workers tracing (telemetry.sdk.name = workers-observability) records no span events, no status description, and has no outcome setter — a custom span can only setAttribute() — so every error span it exported collapsed into a single "Unknown Error" issue per service. That is why the landing worker ships with traces disabled (unchanged here).

The shared error-events SELECT now resolves the exception in this order:

1. the first exception span event — taken verbatim, empty values included, so a span that has one hashes exactly as before;
2. exception.type / exception.message / exception.stacktrace span attributes;
3. semconv error.type + error.message;
4. StatusMessage, then Unknown Error.

The message signature and display label are cut from a new _msgText: StatusMessage whenever it is set or an event exists, and the attribute message only for an event-less span with none. The only rows whose hash changes are the ones that used to share the "Unknown Error" bucket. The 0016 4xx guard keeps dropping a client span whose only error.type is the bare status code (HTTP semconv sets that on any non-2xx response), so the bot-404 noise stays out.

Changes

• packages/domain/src/tinybird/materializations.ts — the MV body; fingerprint.ts gains resolveErrorSource() as the TypeScript mirror of the precedence, with unit tests and SQL-parity assertions.
• packages/domain/src/clickhouse/migrations/0027_error_events_attribute_fallback.ts — drops and recreates both views with the verbatim v27 DDL. requiredForIngest: false; nothing is backfilled (error_events keeps no span attributes, and recomputing FingerprintHash would re-bucket every issue). FINGERPRINT_VERSION is not bumped, for the reason 0018 gives.
• Local chDB schema v16 → v17 via local-schema:bump, with the v16-to-v17 edge (drop the two views, bootstrap v17). Generated schema, local DDL snapshot, insert mappings and Tinybird manifest regenerated.
• apps/api/src/services/warehouse/error-events-attribute-fallback.clickhouse.e2e.test.ts — now-relative seeds through the real migration set: a workers-observability span carrying only error.* attributes labels as TypeError; two bugs of one type split while one bug groups across ids; exception.* attributes beat error.* with the stacktrace feeding TopFrame; an event still wins over attributes; the StatusMessage and Unknown Error fallbacks are intact; the bot-404 stays excluded while a real 4xx exception is kept; error_events_by_time matches byte for byte.
• Docs (warehouse-rollups.md, error-issue-lifecycle.md) and the chat prompt's description of "Unknown Error" updated.

Reviewer notes

• Rows already in error_events keep their Unknown Error label and hash for the rest of their TTL; attribute-only spans land under their real type from cutover forward.
• Tinybird: this changes the SQL of two materialized pipes, so deploying it means altering them; Tinybird CD is currently disabled.
• Verified: domain/cli/api typecheck, domain unit tests (93), CLI local-store migration tests (43), and the new e2e (8) against a local ClickHouse. Not run: the repo-wide suite and the native local-store migration shell probe.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

• Merged main three times, after perf(agent-sessions): detect agent traces on a filtered span index, not a raw scan #692 (migration 0024, local v14), feat(warehouse): read the commit from vcs.ref.head.revision, retire deployment.commit_sha #749 (0025, v15) and feat(agent-sessions): filter by model, agent, tool, environment, id, errors, cost and usage, and sort #738 (0026, v16); this change is migration 0027 and local edge v16 → v17, verified on the merged tree (both ClickHouse e2e suites, domain and CLI tests, domain/cli/api typecheck).