fix(state): stop delegate/tool children corrupting compression lineage

[kshitijk4poor]

Summary

Delegate and tool child sessions no longer corrupt a conversation's compression lineage — session .md export now follows the real continuation instead of a subagent's transcript.

Root cause: get_compression_lineage's forward walk accepted any non-branch child of a compressed session as the compression continuation. A delegate subagent (_delegate_from) or tool-tagged (source="tool") child created before the real continuation row — the common runtime ordering, since the subagent exists before compression rotates the session — was picked as the lineage successor.

Context (real-world impact)

Reproduced on current main: a CLI session that spawns a delegate child and then compresses exports the delegate's private transcript as its continuation:

main:  lineage(root) = ['root', 'delegate']          # subagent hijacks the chain
       lineage(continuation) = ['continuation']      # real continuation orphaned
       export_session_lineage('root') messages: ['root user msg', 'DELEGATE PRIVATE msg']

fixed: lineage(root) = ['root', 'continuation']
       export messages: ['root user msg', 'continuation msg']

Anyone using /export (or anything built on export_session_lineage / get_compression_lineage) on a session that delegated work before compressing gets the wrong transcript today.

Changes

• hermes_state.py: rename _is_branch_child_row → _is_explicit_fork_child_row; treat _delegate_from and source="tool" rows as explicit forks alongside _branched_from; the forward walk now requires _is_compression_child_row instead of merely excluding branches.
• tests/hermes_state/test_session_md_export.py: extend the lineage test with delegate/tool isolation asserts; new regression test for the fork-child-before-continuation ordering including export content (mutation-checked — fails on main, passes with the fix).

Validation

	Result
tests/hermes_state/	49 passed
tests/run_agent/	1438 passed, 1 pre-existing failure (reproduces on clean upstream/main)
Mutation check	new test fails on main's hermes_state.py, passes with fix
ruff	clean

Sliced from #79024 by @RyderFreeman4Logos — cherry-picked with authorship preserved. The prompt-cache-scope portion of that PR is a design question tracked in #79017 (alongside #78956) and is intentionally not included here.