fix(state): stop branch replay at fork snapshot

[QuarkAssistant]

Summary

• walk transcript replay through compression continuations, but stop at the nearest explicit branch edge
• keep presentation lineage unchanged for branch trees and conversation-root resolution
• make resume counting and safety guards use the same branch-aware lineage as the actual resume read

Why this remains needed after #86775

#86775 fixed direct branch resumes on main, but its tip-only branch check still walks through the original parent after a branch is compressed. The result is a duplicated fork snapshot plus post-fork parent turns in the resumed continuation.

This PR detects a branch boundary on each traversed edge. Inherited _branched_from provenance on compression continuations is not treated as a new fork unless it names that row’s immediate parent.

Regression coverage

Real SessionDB tests cover:

• direct branch resume excludes parent turns written after the fork
• nested branch resume excludes parent turns written after the nested fork
• ordinary compression continuation still replays ancestors
• compression after a branch replays from the branch boundary, including inherited branch provenance
• resume counts and size guards match the branch-aware transcript actually materialized

TDD evidence

• current main + the regression file: 3 passed / 1 failed at test_compression_after_branch_replays_from_branch_boundary
• inherited-marker variant before the edge comparison: 3 passed / 1 failed
• resume-count regression before guard alignment: 4 passed / 1 failed (8 != 6)
• final focused file: 5 passed

Validation

• scripts/run_tests.sh tests/hermes_state tests/test_hermes_state.py -q → 375 passed, 0 failed, 2 skipped
• uv run ruff check hermes_state.py tests/hermes_state/test_branch_replay_lineage.py → clean
• python -m py_compile, git diff --check, and ancestry check against current origin/main → clean

Closes #77375

[Enough1122]

AI code review — automated review for reference, author can ignore or act on any point.

fix(state): stop branch replay at fork snapshot

1. hermes_state.py:10479-10500 — _session_replay_lineage_root_to_tip caps the walk at 100 iterations (parity with _session_lineage_root_to_tip, which also caps at 100). The new call sites change the consequences of that cap: get_resume_message_count and assert_resume_safe now count the capped lineage, so a lineage deeper than ~100 compression continuations would undercount and assert_resume_safe could green-light a resume that then exceeds the message cap at materialization time. Pre-existing cap, but now it gates a safety check — worth a comment or a shared constant so the limit is visible in both functions.

2. The boundary condition branch_parent == parent_id (line 10492-10494) depends on _branched_from always naming the immediate parent edge. The compression-inheritance case is well covered by tests, but if any branch-creation path ever writes _branched_from with a non-parent value (or a legacy row records it differently), the boundary is not recognized and live parent turns leak into the branch — the exact bug class this PR fixes. Consider asserting the invariant (marker == parent) at branch creation time, or documenting that the marker must name the fork edge.

3. get_conversation_root still uses the old _session_lineage_root_to_tip full walk, which crosses fork edges — so a branch session's "conversation root" resolves into the live parent chain while replay stops at the fork. If any UI derives breadcrumbs/display lineage from get_conversation_root, it may show the parent turns that resume no longer includes. Confirm this split (presentation lineage vs replay lineage) is intended, or align the two functions.

Verified: _is_explicit_branch_session has no remaining callers after removal, and the new tests (nested branches, compression-after-branch) cover the tricky boundary cases well.