Record skipped-turn telemetry in compatibility runs

[flemming-n-larsen]

Implements CH-018 in the G-001 parity campaign (P-001 / M-006): opt-in capture of skipped-turn callbacks received by bridge bots during compatibility runs.

What this change provides

• When --capture-skipped-turns is enabled, the bridge records distinct (bot ID, round, turn) callback tuples, including warm-up turns, and the harness persists them with the compatibility-run observation.
• Capture states remain distinct: captured (including a verified empty list), disabled, unavailable, and incomplete. Disabled capture removes an inherited telemetry JVM property so it cannot silently enable recording.
• Collection is buffered and flushed at battle end. The frozen robocode.* API and read-only rumble jars are unchanged.

Evidence

• python -m unittest test_parity_registry: 42 tests pass. :robocode-api:test passes, and clue validate --forbid-changes passes (68 artifacts).
• Three forced-skip, one-round runs with two bots recorded 84, 82, and 76 tuples. Every tuple matched exactly one SkippedTurnReported callback line; per-bot completion counts matched.
• An inherited JAVA_TOOL_OPTIONS=-Drobocode.bridge.skippedTurnTelemetry=true did not enable recording when capture was disabled.
• A completed three-member team run reported captured with zero events despite having two runner score groups, confirming completion is checked against staged bots.
• GitHub unit-test and Cliewen-validation checks pass for head 576912114f1d66ffcc96068b9e217a3cbff9d725.

Acceptance limits and remaining work

• Telemetry observes callbacks that reach bridge bots; it cannot report server events that were never delivered to the bridge. The local runs do not establish throughput at 30 TPS or performance on slower machines.
• HARN-008 remains @draft, and M-148 and M-006 remain todo. The reviewer verdict is verifies-something-adjacent: AN-016 records relevant local integration runs, but Python is not yet an accepted evidence carrier and the complete HARN-008 integration scenario is not resolved.
• Two context-isolated review passes found no blocking issues. One advisory remains: no current acceptance criterion owns the classic-version guard covered by four generic unit tests; those tests are not counted as HARN-008 evidence.
• This PR does not publish a release.