Add hint quality probe tool

[austin-smith]

What Changed

Checks in the hint-quality probe harness at tools/hint-probe/ — the acceptance and regression instrument for the hint engines, previously an out-of-tree scratchpad script. Three files: the probe (main.swift), a compile-and-run script (run.sh), and a README that serves as the durable baseline ledger.

Every variant gets identical treatment: 500 seeded deals played to completion by a bot following every hint, alongside a forward-only random control that calibrates each variant's deal universe. All players use the planners' deterministic entry points (no wall-clock deadlines), so every figure is exact and reproduces bit-for-bit across runs and machines. Deals run in parallel across cores (~7 minutes for a full pass; ~40 seconds for a single variant). Acceptance gates are machine-enforced: the probe exits nonzero if any hint follower ever loops, or if Yukon records a single position revisit.

Certified baselines (single run of this exact code, exit 0):

Run	Following every hint	Random (control)
yukon	62.0%	13.6%
klondike draw-1	44.4%	39.4%
klondike draw-3	24.0%	6.0%
freecell	99.8%	0.2%

The README documents how to read these honestly (draw-1 random wins 39.4% because unlimited stock passes reward persistence — hint value there is efficiency, 133 vs 358 median moves), the acceptance checklist for adding a new variant, and the findings from two evaluated-and-retired tap-policy controls.

Why

The recorded win-rate baselines are load-bearing: they are the yardstick for any planner tuning or shared-code refactor, and they are only comparable on an identical deal generator and driver policy. Keeping the harness out-of-tree meant rebuilding it from memory notes — a rebuilt harness that differs subtly invalidates every recorded number silently. Checking it in pins the instrument to the baselines it produced.

Not part of the app target, the test suite, or CI: tools/ sits outside all Xcode synchronized folders, and the probe is a statistical instrument run deliberately, not a pass/fail check.

Validation

• Final certification run of the committed code: exit 0, zero gate violations, output matches the README table verbatim.
• The hint-following column reproduced identically across five consecutive full runs, serial and parallel (parallelism only changes wall-clock time; each seeded game is self-contained and results aggregate in seed order).
• The loop gate is proven live: during development it caught a follower misclassification (FreeCell's nudge-fallback circling on the one solver-unprovable deal) and correctly failed that run's exit code.
• CLI rejects invalid inputs (zero deals, unsupported draw counts) with usage and exit 1.
• App build verified unaffected after adding tools/.