test(ios): establish snapshot convergence baselines and permanent evidence

[thymikee]

Parent: #2188

Outcome

Check in the reproducible evidence needed to judge every later iOS snapshot-convergence change.
A third party must be able to produce current-HEAD local, proxy, compute, and package-size baselines
from documented commands without reconstructing artifacts from an investigation session.

Exact scope

• Add benchmark definitions, a versioned raw-result schema, and one documented reproduction entry
  point under repository tooling rather than a runtime package.
• Measure current runner-backed iOS Simulator snapshots in explicitly named states:
  • cold-cold: Simulator off, daemon off, no prepared runner/bridge cache;
  • cold: Simulator booted, daemon/helper processes absent;
  • warm: Simulator, app, daemon, runner, and target generation ready;
  • relaunch: same Simulator/tooling with a new app process/generation.
• Include representative quiet, list, nested-scroll, alert/system-surface, and XCTest-stress screens.
• Report daemon-resident request time and fresh-process CLI wall time separately.
• Record direct-proxy baselines for persistent-client and fresh-process CLI execution at controlled
  approximately 0 ms, 20 ms, and 80 ms RTT. Record configured bandwidth/packet loss, response
  bytes, median, p95, failures, and raw samples.
• Record current packed, clean-installed, and bundled npm sizes using the repository's existing size
  workflow.
• Promote iOS: skip private-AX deep extension on observation-only captures (settle, stabilization, corroboration) #1626's deep-button false-failure into permanent, implementation-independent fixture
  evidence: a shallow surface observation must not prove that a full action had no effect when the
  changed descendant was omitted.
• Link App-mount race on fast/cold opens: first capture can see an unreadable or empty tree (cross-platform) #1571 in the result schema and classify an unreadable/empty first tree separately from bridge,
  runner, timeout, and stale-generation failures.

Use at least 20 samples per warm cell and 10 per cold cell unless the checked-in harness documents a
larger statistically justified minimum. Store raw samples; summaries alone are insufficient.

Non-goals

• No production routing, snapshot semantics, bridge, runner-demand, or fallback change.
• Do not implement iOS: skip private-AX deep extension on observation-only captures (settle, stabilization, corroboration) #1626 or expose acquisition intent publicly.
• Do not claim the investigation's provisional numbers as reproduced until this issue's commands
  produce them on current HEAD.
• No physical-iPhone measurement; that track is separate.

Acceptance and evidence

• One documented command reproduces each local measurement cell and writes schema-valid raw output.
• One documented command reproduces each proxy cell without changing application semantics.
• Repeated runs identify machine/toolchain, package revision, screen fixture, target generation,
  sample count, failures, and outliers.
• The deep-button fixture is observed failing under the invalid shallow-corroboration rule and green
  under the current safe/full requirement. Record the exact planted-red command and failing assertion.
• Package-size output reports packed, installed, and bundled values from the same revision.
• Focused tooling/schema tests and pnpm check:affected --run pass. Native and proxy measurements
  are attached as evidence rather than represented by fixture-only tests.

Worker stop conditions

• Stop rather than substitute ad-hoc timing if a required state cannot be produced deterministically.
• Report infrastructure contention separately; do not convert changing timeout sets into product
  regressions without an origin/main control under the same load.
• Do not modify or close iOS: skip private-AX deep extension on observation-only captures (settle, stabilization, corroboration) #1626 or App-mount race on fast/cold opens: first capture can see an unreadable or empty tree (cross-platform) #1571.