fix(core): share child prompt cache affinity

[rekram1-node]

Summary

• use the parent session ID as the prompt cache affinity key for child sessions
• preserve existing top-level and fork cache affinity behavior
• cover child runner and compaction requests

Testing

• bun typecheck (packages/core)
• focused session runner cache-affinity tests
• bun test test/session-compaction.test.ts
• bun run check

[rekram1-node]

Validated with real OpenAI gpt-4.1-mini sessions in an isolated A/B against the pre-PR commit.

Fixture:

• empty temporary project and dedicated SQLite database
• project config discovery, MCPs, and unrelated tools disabled
• three fresh foreground general children under one parent
• identical model, system prompt, and user prompt
• children created sequentially with sessionID omitted
• static prompt deliberately above OpenAI's 1,024-token cache threshold

OpenAI-reported child usage:

Build	Child 1 cache read	Child 2 cache read	Child 3 cache read
pre-PR	0	0	0
PR	0	4,992	4,992

Each child request contained about 6,319 prompt tokens. On the PR, warmed siblings reported about 1,324 uncached input tokens plus 4,992 cached tokens; before the PR every sibling reported the full ~6,319 as uncached. This confirms that parent-derived affinity produces real cross-child cache reuse.

The remaining uncached ~1.3k tokens are consistent with the separate early session-ID/system-context prefix divergence and are intentionally outside this PR.