fix: the maximal create runs once the minimal object is gone

[ShocOne]

No entity ever recorded a maximal body, so every acceptance configuration fell back to values derived from the document — the thing #119 exists to stop.

The cause

The strategy program created the widest valid body while the object the minimal create had just made was still live, and both bodies are synthesised from the same document. An API that keys an object on any field they share answers the second create with a conflict:

attempt 0  400  "Invalid icon value: tfpfgen-test-icon"  -> drop icon
attempt 1  409  conflict
attempt 2  409  conflict
...                                                       -> to the allowance, then give up

The reduction was working the whole time. It dropped what the API named, and then had nothing left to learn from — a conflict says nothing about how wide a valid body is. buildProgram put the teardown last, unlike the non-strategy plan in plan/steps.go, which deletes before the maximal create.

runCreateMaximal makes a second object and deletes it again, so it belongs after the delete rather than before it.

The result

Maximal bodies are recorded, and the maximal configuration is now genuinely wider than the minimal one — and excludes what the API accepts without ever returning:

resource "thousandeyes_tag" "test" {
  access_type = "all"
  color       = "#FF0000"
  description = "To tag assets in San Francisco"
  key         = "branch"
  object_type = "test"
  type        = "static"
  value       = "sfo"
  # match_type skipped: the API accepted this property and did not return it, so terraform cannot hold it in state
  # filters skipped: the API accepted this property and did not return it, so terraform cannot hold it in state
}

TestUnit_Strategy_TheMaximalCreateFollowsTheDelete pins the ordering.

Also here

NamesRefusedFieldInProse on the quirkserver, with an exhibit. It could only model the unhealable refusal shape — a bare field name in detail beside a generic title — so no test could exercise a create the adjustment loop heals. Real APIs write both forms; now so does the quirkserver.

The recipe learns the body that worked rather than keeping the plan's document-derived guess. Everything downstream replays it. No test here demonstrates a changed outcome — each of those paths heals again on its own, so the guess survives being replayed — and the commit says so. It is still the wrong thing to carry.

handoff.md rewritten. The previous version opened by saying no audit had ever run against any pilot. It now carries the full test-harness reference: which script is the benchmark, what the loop script does, where the credential lives, which files each generated resource carries and which are replayed rather than derived, where the audit writes what it learned, and the ranked next steps.

Measured

sh tf_acceptance_tests.sh — 1 passed, 34 failed, from 0/35 at the start of this work. 33 of the 34 fail at step 1, because only four entities have a recorded body; the rest still have document-derived configs. Getting more entities through the probe is the next lever and is the top item in the handoff.

make check green at 90.7%.

🤖 Generated with Claude Code