Make Parallel Sub-Agents Share the Prompt Cache

PROJECT CONTEXT:
My parallel sub-agents each pay full input-token price on every request, making fan-out prohibitively expensive. Claude Code's fork mode exists to fix exactly this: all engineering choices - placeholder results, exact tool arrays, inherited configs - serve making sibling request prefixes byte-identical so they share one cache entry.

GOAL:
Engineer fork dispatch for maximum prompt-cache reuse: stable prefixes, identical placeholders, exact configuration inheritance, and measured cache-hit rates.

REQUIREMENTS:
- Sibling fork requests identical in every byte up to the per-child directive
- Placeholder tool-results with fixed, shared text for pending parent tool calls
- Tool definitions passed through exactly - no re-filtering, re-ordering, or re-serialization
- Model-configuration inheritance (thinking/verbosity settings) from the parent
- Cache-hit metrics per dispatch batch proving the savings are real

CONSTRAINTS:
- Any 'harmless' prefix difference (timestamp, ordering, whitespace) silently zeroes the savings
- The differing content (the directive) must come last in the request

DEFINITION OF DONE:
- A byte-level diff of two sibling requests shows differences only in the final directive block
- Provider metrics show cached input tokens dominating for second-and-later siblings
- Fan-out cost scales near-flat with sibling count instead of linearly
- A prefix-stability regression test guards the dispatch path

COMMON FAILURES TO AVOID:
- Per-child placeholder text ('processing task 1 of 3') breaking prefix identity
- Re-running tool filtering for each child, changing array order and killing the cache
- Injecting the child directive before the shared history instead of after
- Assuming savings without measuring - one serialization change and you pay full price silently