docs: weight the test tiers for the one user who runs this

[benjipeng]

Documents only. No code changes.

Why

The testing standard described six tiers without saying what they are worth here. Plexmaton has
one user, who runs it daily and reports what breaks — so the usual reason to invest in end-to-end
automation is already covered, at a price the corpus was paying without saying so.

That price became visible while fixing compaction: provisioning_process_death_recovers_one_exact_passive_child
spawns and kills a real OS process from inside cargo test, and fails three runs in four on
unmodified main. Four documents cite it as proof of five invariants. A test that mostly fails is
worse than no test, because the corpus reads as settled when it is not.

What changed

standards/testing.md

• Tiers 1–3 are the instrument, because they shape behaviour. A Tier 4–5 test earns its place
  only by reaching what no assertion can: terminal lifecycle, escape sequences, a real process
  boundary. Anything else buys detection the user already gives.
• A settled boundary is a precondition for Tiers 4–5, not for Tier 1. Much of this project has
  boundaries nobody has drawn. A unit test is welcome there, and mocking the edges is often how
  the boundary gets drawn at all; an end-to-end test has nothing to assert against a system nobody
  has decided yet, and repairing one defends a guess.
• A Tier 4 test belongs in its own lane. One that spawns or kills a process while sitting in the
  fast suite charges the whole development loop for evidence only a wrap-up ever needed.
• The flaky rule is rewritten: a flaky test proves nothing and is worse than none while a document
  still cites it. Fix it, or move it to a reporting lane and mark what it claimed unproven.

.agents/README.md — an evidence row says what kind of proof it is. A fast deterministic test,
a rendered frame the user has read, and a real-terminal script all prove something and do not prove
it equally; listing them alike leaves a reader unable to tell how firmly an invariant stands.

phases/phase-04-product-polish.md — stage 29 recorded complete (the compaction decline, #35),
and two unstarted stages record what the code does not yet follow:

• 30, automatic compaction is invisible. A checkpoint the runtime takes on its own reports
  nothing, while the same operation asked for by name reports a line. Needs a rendered frame
  first: an automatic checkpoint lands mid-turn, where a requested one lands after the last entry.
• 31, test evidence says what kind it is. The process-spawning test's lane, the 46 smoke-script
  citations across 23 evidence documents, and this standard outgrowing its 8 KB budget while it
  carries the change.

Note on the budget warning

standards/testing.md was 8186 of 8192 bytes before this change — already full. It is now 931
over, and .agents/README.md 389 over. Both warnings are advisory. Part of the excess is
transitional: the paragraphs explaining why the policy is changing become unnecessary once the
code follows it, and the stage-31 notes are deleted when consumed. The rest is genuinely new
policy, and stage 31 records that the standard's Owns lists four things while performance
evidence is a separate trigger.

Citations and frames gates pass; six pre-commit gates pass.