feat(web): wire compaction tab to the real assistant route

[vellum-apollo-bot]

Why

The Compaction tab was showing compactions outside the window between the previous LLM call and the selected one — three events frozen at 11:32 / 11:54 / 12:18 PM, regardless of which call you clicked.

The screenshot Vargas surfaced this morning is the mock: compaction-trail-mock.ts hardcodes MOCK_EVENTS at Date.parse("2026-05-26T15:32:11Z" / "T15:54:38Z" / "T16:18:02Z") and slices the first N based on a hash of the callId. There's no relationship between those timestamps and the rail.

The real route was already shipped and already implements the right semantic — assistant/src/runtime/routes/conversation-compaction-routes.ts resolves the floor to the most recent non-compactionAgent call before the selected one and returns compactions strictly in (floor, selectedCall.createdAt). The frontend just never got wired to it.

What

• compaction-trail-fetch.ts — new fetchCompactionTrail(assistantId, conversationId, callId, signal). Hand-rolled because the route isn't in the generated SDK yet; mirrors fetchConversationMessages and archiveConversation (path params via SDK options, no trailing slash on the action segment).
• compaction-trail-fetch.test.ts — five tests, spying on client.get per the messages.test.ts precedent. Pins URL pattern, path/query params, abort-signal forwarding, HTTP-error → CompactionTrailRequestError, malformed-body branch.
• compaction-trail-api.ts — swap the mock import for the real fetcher. Query shape unchanged, so the Compaction tab itself needs no edits.
• compaction-tab.tsx — doc comment updated to describe the route semantic instead of the mock.
• compaction-trail-mock.{ts,test.ts} deleted (the mock's own doc-comment said to do this when the real route shipped).

Verification

Simulated the real route against the actual 81cc6bf4 conversation that surfaced the original investigation:

Selected call	Prev real call	Compactions in window
post-comp#1 (22:19:20)	22:18:05	1
post-comp#2 (22:20:38)	22:19:20	1
post-comp#3 (22:28:49)	22:26:54	1
post-#4,#5 (22:31:34)	22:28:49	2
post-comp#6 (22:34:05)	22:33:03	1

Tight prev-call → selected-call windows, never more than 2 (and that's the back-to-back emergency-then-retry pair). Each call in the rail now shows only the compactions that built its own context.

What this PR does not do

• Doesn't surface the silently-dropped-vs-actually-applied signal — that's still a separate workstream item (the visualization that distinguishes the two devastating bugs from the original investigation).
• Doesn't add durationMs — llm_request_logs doesn't carry it; the route ships null and the UI renders "Unavailable". That gap stays visible until we decide whether to extend the column or stand up a compaction_logs table.

Test plan

• bun test src/domains/chat/inspector/compaction-trail-fetch.test.ts (5 / 5 pass)
• bun test src/domains/chat/inspector/ (47 / 47 pass)
• bun run typecheck clean
• bun run lint clean