feat(rate-of-closure): bind optimizer to full flight

[d-sorgcodexagent]

Summary

• bind player-capability-profile/v1 and capability requests to the real Waterloo/Penner forward-flight model in Python and React
• validate units, clubs, safe bounds, physical domains, sampling, target-frame signs, and explicit per-club spin provenance before integration
• return the complete available ball-flight-result/v1 scalar set with typed complete/nonconverged/failed outcomes
• centralize gyro-projected spin-axis tilt across result, impact, and variation producers
• add a shared 16-metric parity fixture and update the root, campaign, Rate, specification, and contract handoffs

Contract decisions

• no hidden global driver spin fallback; three-variable profiles require a sourced spin default for every requested club
• positive spin-axis tilt means fade/right in the target frame
• Python records adaptive SciPy RK45 and React records fixed-step RK4; parity uses published tolerance bands, not bitwise identity
• this post-impact evaluator cannot fabricate no_impact

Validation

• Python shared flight suite: 138 passed, 4 optional-Rust skips
• React suite: 97 files, 597 tests passed
• Ruff check and format, targeted mypy, TypeScript type-check, zero-warning ESLint: passed
• Vite production build: passed, 176 modules transformed
• independent final review: no P0/P1/P2 blockers
• staged diff/secret pattern/structural budget checks: passed

Stack

This draft intentionally targets feat/4197-capability-observer at 49612946138b1021f80c9f8d2a4d06f1610825db (PR #4283). Do not retarget or merge it out of order.

Progresses #4197. The issue remains open for the PyQt6/React authoring, persistence, worker, observation visualization, and release work.

[dieterolson]

Defect found from the stacked child (#4294): the TypeScript evaluator throws untyped exceptions on ~12.5% of a default driver search.

While verifying the Shot Optimizer UI against this evaluator, the React runtime reported 12 of 96 samples as evaluator_exception where Python reported zero. Deterministic reproduction against the default workflow document:

Python  : complete 72 | failed 24 (all no_ground_crossing_before_max_time)
React   : complete 66 | failed 18 (no_ground_crossing_before_max_time) | failed 12 (evaluator_exception)

Root cause, from wrapping the evaluator to un-suppress the message:

RangeError: timeS must be nonnegative
  at validateInputs (src/rate_of_closure/web/src/model/ballFlightMetrics.ts:110)
  sample = {ball_speed: 83.24, launch_angle: -4.32, launch_direction: 9.21}
  sample = {ball_speed: 85.84, launch_angle: -3.17, launch_direction: 10.25}

Every failing sample has a negative launch angle at high ball speed — legal inside the profile's declared safe bounds (launch_angle spans -10°..45°), so the search reaches them normally. The Python path handles the same samples without throwing.

Why this matters beyond the count:

• It contradicts this slice's own spec text: "Metric sets, frames, signs, and typed statuses are parity contracts" — the same physical sample yields a typed status in Python and an untyped exception in TypeScript.
• It also contradicts "Contract and programming errors surface instead of being silently counted as ordinary failed trials" — a RangeError from a validator is precisely a contract/programming error, and it is currently being absorbed into the ordinary failed cohort.
• ballFlightMetrics.ts was last touched by c280407d4 (this PR's head), so it belongs to this slice rather than the observer parent or the UI child.

Not fixing it from the child branch — that would pull a parent-owned file into #4294's diff and cross the stack. Flagging it here so this slice can decide whether the trajectory builder should not be emitting a negative timeS for descending-launch cases, or whether the validator's precondition is wrong.

Reproduction note: the 12 exceptions are invisible in the compact result, which only exposes failed_count. They became visible because #4294 now prints the retained reason_code values alongside the counts.