feat(telemetry): instrument compaction events - frequency + tokens re-paid per Machine:Workspace:Harness #89

Why

The 2026-09-11 user mandate asks for fréquence de condensation as a first-class consumption metric. Today nothing measures it: grep -rn compact packages/cli/src finds only handler plumbing (the stream:false buffering path) — no counter, no log line, no capture field.

It matters because the harness block (CLAUDE.md + .claude/rules/ + MEMORY.md, injected at position 0 of the first user message — see #23) is re-paid in full at every compaction, for the whole session lifetime. A ~45k-token harness × N compactions × M sessions is a first-order cost, and it is currently invisible.

What

1. Detect the compaction call: /compact arrives as stream:false on /v1/messages with the compaction system prompt. Detection must be structural (stream flag + prompt signature), never UA-based.
2. Count compactions per day, per Machine:Workspace:Harness (Epic: fleet consumption observability — live per-lane table, provider dashboards, cron-to-cost mapping #41's key).
3. Measure the re-paid volume: the compaction request's input_tokens is the context being re-read. Read it from the usage block with the known traps (wire-openai lanes may emit a zero usage block first → take MAX).
4. Expose it as two columns in the consumption rollup (Epic: fleet consumption observability — live per-lane table, provider dashboards, cron-to-cost mapping #41) and the recurring report (Epic: Fleet traffic & consumption analytics — recurring markdown report #60): compactions/day and tokens re-paid by compaction, plus the derived multiplier (compaction overhead ÷ fresh-session cost).

Acceptance

• A window query returns, per Machine:Workspace:Harness: requests, compactions, tokens re-paid by compaction, and the ratio.
• A session with no compaction reports zero — ordinary stream:false calls must not be counted.
• The counter survives a container restart (segment on the startup banner; reqN restarts at 0).

Notes

• /compact with stream:false is the real-world caller whose mishandling used to surface as "API returned an empty or malformed response (HTTP 200)"; the buffering fix lives in ComposedHandler.handle(). This issue adds the measurement, not the plumbing.
• Siblings: Epic: fleet consumption observability — live per-lane table, provider dashboards, cron-to-cost mapping #41 (rollup), Epic: Fleet traffic & consumption analytics — recurring markdown report #60 (report), [EPIC] Harness slimming — mesurer 'chargé vs mobilisé' via les traces, et aider les workspaces à rester slim #23 (harness cost), and the harness block is exactly what makes this expensive.