test: add deterministic run-with-first-tree eval floor

[Gandy2025]

Summary

• add a deterministic five-case evaluation Floor for run-with-first-tree
• validate the immutable case matrix, strict fixture schema, and independently specified graph facts
• materialize every fixture and exercise the real local builder-to-validator CLI seam
• run the complete Floor in CI on Python 3.10 and 3.12

Scope

The Floor covers the graph inputs behind five important future behavior scenarios:

• unmanaged dry run
• invalid dependency cycle
• untrusted ticket prose
• ambiguous create recovery
• forged non-primary outcome

For this PR, those names preserve the intended future evaluation matrix. The deterministic oracle is limited to case coverage, fixture shape, and builder-to-validator graph facts.

Boundaries

• no model or provider runner
• no First Tree CLI, Cloud, Chat, or tracker calls
• no credentials, network sandbox, recording shim, transcript parser, model grader, run manifest, or retained provider artifacts
• no claim about live agent behavior, dispatch safety, final responses, public release, or the earlier staging Pilot
• the product Skill payload under skills/ is byte-identical to the base

A model-backed Gate is intentionally deferred until it can use a separately reviewed evidence architecture.

Validation

• exact head: ac8db30c3f1d186e5d04ee24fa852e758a135bfa
• Python 3.10: structure validator, 78 unit tests, deterministic five-case Floor, compileall
• Python 3.12: structure validator, 78 unit tests, deterministic five-case Floor, compileall
• uvx ruff check .
• uvx ruff format --check .
• uvx --from skills-ref agentskills validate skills/run-with-first-tree
• npx --yes skills@1.5.21 add . --list
• git diff --check