feat(logbuf): Cisco-style "last message repeated N times" dedup

[arcaven]

Summary

Closes the symptom Skippy reported in aae-orc-1d2: a single "ssh: client connected" line at 0.5 Hz dominated the daemon's 10k-line ring buffer, shrinking the visible-history window from days to ~5.5 hours. Cisco-style "last message repeated N times" dedup collapses adjacent identical lines into one stored slot plus a counter — a 1000-line run consumes 2 ring slots, not 1000.

Behavior

Before:                       After:
ssh: client connected         ssh: client connected
ssh: client connected         last message repeated 999 times so far
ssh: client connected         (next interesting log line)
... (998 more) ...
(next interesting log line)

When a different line arrives, the buffer emits a last message repeated N times summary before recording the new line. Tail() synthesizes a ... so far summary for any still-active run so an operator querying mid-burst sees it without waiting. Flush() is the periodic-cut hook for runs that never see a different line; not wired to a timer in this PR (caller decides).

Design choices

Captured in _kos/ideas/log-rrd-deduplication.md (the aae-orc-4wz brainstorm):

• Previous-line-only matching (Cisco original). A B A B does not collapse — documented in TestDedup_NonAdjacentNotDeduped.
• No regex-based key normalization in this pass. Lines must match exactly. The three motivating cases (ssh poll, crashloop chatter, future debug spam) are all monotonic, so exact-match catches them all without the false-collapse risk of regex normalization.
• Synthesized active-run summary in Tail() rather than a stored "live" summary that would have to be replaced on each new repeat. Keeps the stored buffer clean and Flush() simple.
• Flush is idempotent so a periodic timer wiring is safe with no was-active bookkeeping at the call site.

Test plan

• Existing tests still pass (TestWrite_LargeSinglePayload updated — 100 identical writes now stored as 1 line + active count, was 3 lines bound to cap)
• TestDedup_RunBrokenBySummary — A A A B → [A, "repeated 2 times", B]
• TestDedup_NonAdjacentNotDeduped — A B A B → no collapse
• TestDedup_TailSynthesizesActiveSummary — A A A A → [A, "repeated 3 times so far"]
• TestDedup_FlushEmitsSummary — Flush appends summary and clears active run
• TestDedup_FlushIdempotent — second Flush is no-op
• TestDedup_SingleOccurrenceNoSummary — A B → no spurious "repeated 0 times"
• TestConcurrentWrites still passes with -race
• golangci-lint clean

Related deferred work

aae-orc-4wz (the broader RRD-style probe with key extraction, structured repeat_count in the logs RPC result, OTEL bridge) stays open. This PR ships the smallest version that makes the concrete 1d2 symptom go away.

Refs: aae-orc-1d2, aae-orc-4wz