fix(sdk_dispatch): refuse scalar silence-threshold on all phases

[oprince]

Summary

• Broaden the postmortem fix Agentic Systems Experimentation and Discovery (ASED) #2 (which refused bundle scalar recommended_turn_silence_threshold_seconds on design only) to refuse it on all three phases: design, execute_analyze, report.
• Only the per-phase map form is honored now. An explicit {design, execute_analyze, report} entry means the author intentionally tuned that phase.
• Motivated by an IOCR follow-up run (iocr-hrl-ocr-models-detection-word-unet-v2-cand-0001) that reproduced the same class of failure on execute_analyze instead of design: iter-2 rehearsal measured per-image OCR latency (~3.5s) and emitted a scalar 15s threshold, but iter-3 execute_analyze wraps a 10+ minute subprocess (6 policies × 25 repeats × 8MP-image benchmark). The SDK-event gap was 621.4s, so the 15s watchdog fired on every attempt and burned 71 minutes silently.
• The scalar's assumption "portable across phases" doesn't hold: every dispatched phase wraps LLM-driven tool calls around a benchmark, and subprocess wall-time (not per-op timing) is what the watchdog actually sees. The bundle writer emitting the map form remains the durable upstream fix (postmortem fix Agentic Strategy Evolution: a three-loop methodology for optimizing multi-layer policy spaces #1).

Test plan

• `pytest tests/test_experiment_spec.py::TestWatchdogReadsBundleOverride -v` (13 tests pin the resolver contract in both directions: scalar refused for every phase; map honored for every phase)
• The two existing override-and-restore tests were converted to the map form — confirm their invariant is still covered
• Full test suite: `pytest tests/test_experiment_spec.py`

Closes #305