M3-5: run arm 2 (three-way agent cost comparison) — closes M3

[DeDuva]

Summary

Runs M3-5 arm 2 — the last piece of M3 — and closes the milestone.

Arm 2 measures tokens, tool calls, error rate, and wall clock for the same
task suite completed three ways: GitHub + gh (real GitHub repo,
DeDuva/adp-bench-arm2-scratch), ADP via gh (the compat plane,
unmodified gh through GH_HOST — the MVP's own success criterion), and
ADP-MCP (the native plane's own MCP tool surface,
server/src/mcp/server.ts, run as a real subprocess the agent talks to).

• bench/arms/three-way-cost.mjs: one trial (method × task × rep) per
  invocation, same granularity as arm 1's driver. The real agent is the
  claude CLI, non-interactive, tool-scoped per method via
  --allowedTools/--mcp-config.
• bench/tasks/arm2/: two small, deterministic-to-grade tasks (clamp,
  titlecase) — new to this arm, not reused from squad-lab, so ADP's own
  bench owns its own task suite per the M3-5 layout.
• bench/report/three-way-cost.md: generated from bench/runs/, same
  reproducibility contract as arm 1 — npm run report (now regenerates both
  reports) reproduces it byte-for-byte from committed records.

Result

Pilot slice (12 trials: 3 methods × 2 tasks × 2 reps): 12/12 landed, $1.25
total spend. adp-gh costs about the same as github-gh ($0.0848 vs $0.0850
avg/trial) — once GH_HOST-vs-git-remote host matching and this instance's
one_approval land policy were handled, unmodified gh pays no real
agent-cost tax against ADP versus GitHub itself. adp-mcp costs ~1.7x more
($0.1435 avg/trial), traced to a specific, closeable gap: the ~8-tool native
MCP surface has no proposal-open tool, so an agent using it pays a REST round
trip gh bundles into one command — a finding about today's MCP tool
coverage, not a verdict on the native plane's design. Full writeup and every
number's derivation: bench/report/three-way-cost.md.

Closes M3

All three benchmark arms are now published:

• Arm 1 (merge contention) — deterministic, CI-enforced, unchanged by this PR.
• Arm 2 (this PR) — pilot scale, 12 trials.
• Arm 3 (fan-out vs serial) — run in squad's duva-bench track:
  squad PR #119.

ROADMAP.md: M3 → complete 2026-08-10, Now/Next moved to M4 (not started),
the M3-5 budget item removed from Blockers. docs/pragmatic_mvp.md: the
M3-5 "not run" sentence replaced with the landed status and links, no other
narrative changes.

Scope note

The one implementation choice worth recording: the native MCP tool surface
has candidate-set open/select/resolve but no tool to open the underlying
proposal itself (that verb only exists on the compat-plane REST route,
/repos/{owner}/{repo}/pulls). The adp-mcp arm's agent is given one scoped
curl for exactly that one call — everything else (workspace/branch via
git, landing via the native MCP tools) goes through the native plane. This
is called out in the report's own "what this shows" section as a real
finding, not smoothed over.

Test plan

• make check green (had to stop a leftover long-running dev server
  from earlier in this session that was sharing the test Postgres with
  make test-all and intermittently failing the mirror-mode tests — not a
  regression from this diff, confirmed by re-running test/e2e-mirror.test.ts
  in isolation both before and after)
• npm run report --prefix bench regenerates both reports byte-identical
  to what's committed (git diff --exit-code bench/report/ clean)
• All 12 run records in bench/runs/ reconcile with the numbers in
  bench/report/three-way-cost.md