perf(startup): scope orphan reapers to server commands, share one /proc walk, allow opt-out

[debugmcpdev]

Problem

The two startup orphan reapers (reapOrphanJvms, reapOrphanProxies) are awaited at the very top of main(), before commander even parses argv:

• src/index.ts:108-111 — await Promise.allSettled([reapOrphanJvms(...), reapOrphanProxies(...)])
• createCLI() / program.parseAsync() run strictly after

Consequences:

1. They run for every invocation, including --version, --help, and bad-arg errors — not just when a server is actually starting. A --version call on a busy Linux host pays the full double /proc walk.
2. They block startup by their full duration (~250 ms post-perf(reaper): bound /proc scan concurrency at startup #395 on a 1600-process host; more on darwin/win32 where they shell out to ps/PowerShell).
3. The two reapers walk /proc independently and concurrently — the same readdir('/proc') + per-pid cmdline read happens twice at the same time (2 × PROC_SCAN_CONCURRENCY reads in flight post-perf(reaper): bound /proc scan concurrency at startup #395), for two different marker greps over identical data.

There is also no env opt-out for containerized/ephemeral environments where orphans are impossible (fresh PID namespace) and the scan is pure waste.

Proposal

• Move the reaper invocation into the server-starting command actions (stdio/http/sse) so --version/--help never pay it.
• Share one /proc walk: a single scan produces (pid, argv) pairs; each reaper contributes a matcher over that list. Halves the I/O and the transient allocations.
• Consider firing the reapers after the transport is up (fire-and-forget with logging) instead of blocking startup — orphan reaping is not a precondition for serving requests. Needs care re: process-lifetime races, so measure first.
• Optional MCP_SKIP_ORPHAN_REAPERS=1 escape hatch for PID-namespaced containers.
• Further slimming candidate, measured by the perf(reaper): bound /proc scan concurrency at startup #395 author but deferred to keep that PR's I/O model unchanged: a synchronous openSync/readSync scan with one recycled buffer (their numbers: 57 MB / 25 ms vs 69 MB / 99 ms for the bounded-async version). Requires rewriting the fs/promises mocks in the reaper tests.

Measuring

scripts/mem-bench.mjs (busy-proc scenario) synthesizes an N-process /proc inside the docker image and reports RSS + peak (VmHWM) medians across trials — use it to validate any change here. Baseline numbers are in the PR that added the harness.

Follow-up to #395.