feat(replay): bulk message replay / DLQ triage (UI, REST, CLI, MCP)

[ritikbansod]

Implements F3 — bulk message replay / DLQ triage from the deep analysis, fulfilling the README roadmap's "Message replay: copy a filtered set of messages to another topic or cluster". Stacked on #4 — merge #1 → #2 → #3 → #4 → this.

What it does

POST /api/clusters/{clusterId}/topics/{topic}/replay selects messages from a source topic (typically a DLQ) with the browser's read-only scan — keyContains/valueContains filters, oldest first — and produces them to a target topic on the same or a different connected cluster, preserving keys, headers and original timestamps (so offset-reset semantics stay meaningful after replay).

• dryRun: true previews the selection without producing — the recommended first step.
• Bounded: max 1000 messages per call; reachedEnd tells you whether the whole source was covered.
• Fail-fast on the first produce error, with the source partition@offset in the error list, so a re-run with a narrowed filter can continue.

Safety integration is free — replay is a real mutation, so the chain from #2/#3 applies automatically: admin role only in authenticated modes, every call lands in the mutation audit log, refused outright in KVIEW_READONLY=true mode. The MCP tool additionally carries the same confirm=true guard as topic delete (replaying duplicates messages — the AI is told to dryRun first).

Surfaces

• Web UI: Replay tab in the data explorer — target topic/cluster, filters, limit, headers toggle, dry-run checkbox, outcome summary.
• CLI: kview replay <sourceTopic> --to-topic T [--to-cluster ID] [--limit N] [--dry-run] [--no-headers].
• MCP: kview_replay_messages — 19 tools now.

Bonus fix

MessageBrowserService.browse never stopped early when the topic was smaller than the limit — it polled until the deadline (up to 60s). It now stops as soon as every assigned partition reaches its end offset; small topics answer immediately, and reachedEnd is exact. (The browser test suite dropped from 34s to 15s.)

Verification

• 3 new integration tests against an embedded broker: same-cluster copy (20/20, header retry=3 and original timestamp preserved, all keys present), dry-run selects 4 poison messages without producing then the real filtered run copies exactly those, cross-connection replay through a second stored profile read back via that connection's own handle.
• Full suite: 56/56 green; CLI and MCP smokes pass (19 tools; replay refuses without confirm=true).
• Not covered: binary (non-UTF-8) header values are transported as text — documented limitation; headers are re-encoded UTF-8 on the target.