Efficiency watcher: network bytes and cache hit rates by cause, filed when a cost doubles #7558

Part of the epic #7553. Blocked by the maintainer framework (#7554).

Problem

The npm re-download (~30 GiB/day, #7541) went unnoticed until the steward measured network bytes and read caches by hand. The forge has no cost-per-cause view. #7546 adds the first number (registry bytes per hour); the rest is a watcher that reads it, compares to a baseline, and files when something doubles.

Build (duty: efficiency-watcher)

• Hourly samples by cause: npm registry bytes, image pulls, git fetch/clone bytes, bytes per check/gate/preview run; cache hit rate per cache (bun install cache, check impact cache); install downloads per run; CPU-seconds and peak memory per run (reuse the 15 s sampler, src/boards/sampler.ts:1).
• Baseline: median of the same hour over the last 7 days. Fire when a cost is 2x its baseline for 2 hours, or crosses an absolute ceiling.
• A finding names the cause, the number, the baseline and the code path that starts the cost; one issue per cause, a comment while it stands, closed with a note when it has been back under for 24 h.
• Samples land in a board source so the board can chart them.

Done when

• With a seeded series where npm bytes jump 10x for 2 h, one issue is filed with cause, numbers and baseline; with the series flat, nothing is filed.
• Replaying the #7541 series, the watcher files within 3 h of the first bad hour.
• Tests: test/efficiency-baseline.test.ts, test/efficiency-double.test.ts, test/efficiency-dedupe.test.ts, test/efficiency-close-when-clear.test.ts.
• Depends on the bytes number from #7546 (open; link, do not rebuild). Related: #7541, #7543, #7544, #7545.

Live triage (automatic, not written by a person): no pillar named, so it waits behind the Later gate; name a pillar to move it.