fix(compression): preserve goal, platform, and session indexing across rotation

[teknium1]

Summary

Three pieces of state that used to be lost or corrupted when context compression rotates session_id to a continuation child now survive the boundary. They're fixed together because all three live in the same ~80-line rotation block in agent/conversation_compression.py — separate PRs would cascade-conflict in that exact region.

What changed

• Persistent /goal is lost after context compression rotates session_id #33618 — /goal survives rotation. load_goal does a flat per-session lookup with no lineage walk, so an active goal silently died when compression minted a fresh child id. Added migrate_goal_to_session() (hermes_cli/goals.py) — move-not-copy: copies the goal to the child and archives the parent row as cleared so exactly one active goal row exists. Called after the child session is created.
• Context compression creates orphan sessions missing from state.db #33906 / Context compression creates orphan sessions missing from state.db #33907 — no more orphan sessions. If the child create_session raised (FK constraint, contended write), the outer handler only warned and let the agent continue on the NEW id — which has no row in state.db — producing an orphan (compression_parents_with_no_child in the wild). Now the rotation rolls agent.session_id back to the still-indexed parent (reopening it) instead of stranding the conversation on a phantom id.
• bug: compression boundary drops platform kwarg, causing source=unknown on all subsequent LCM messages #27633 — source=unknown fixed. The compaction-boundary on_session_start notification omitted the platform kwarg, so context-engine plugins (e.g. hermes-lcm) saw source=unknown for every message after the boundary. Now forwards platform= (matching the initial session-start call in agent_init.py).

Validation

Bug	Guard test	Verified
#33618	test_goal_follows_compression_rotation + 5 migrate_goal_to_session unit tests	✓
#33906	test_rolls_back_to_parent_when_child_create_fails	fails without fix → passes with
#27633	test_on_session_start_receives_platform	✓

Real SessionDB + real compress_context integration tests (tests/agent/test_compression_rotation_state.py). 161 passing across the compression + goals suites; orphan-rollback confirmed as a true regression guard.

Credit

Combines the substantive fixes from three contributor PRs into one conflict-free rotation-block change, authorship preserved via co-author trailers: #33620 (denisqq, goal migration), #34048 (zccyman, orphan rollback), #42679 (liuhao1024, platform forward). Those PRs will be closed pointing here.

Infographic

[github-actions]

🔎 Lint report: fix/compression-rotation-state-loss vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 11278 on HEAD, 11276 on base (🆕 +2)

🆕 New issues (2):

Rule	Count
unresolved-attribute	2

First entries

tests/agent/test_compression_rotation_state.py:56: [unresolved-attribute] unresolved-attribute: Unresolved attribute `context_compressor` on type `AIAgent`
tests/hermes_cli/test_goals.py:576: [unresolved-attribute] unresolved-attribute: Attribute `goal` is not defined on `None` in union `GoalState | None`

✅ Fixed issues: none

Unchanged: 5913 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.

[PaddyPatPat]

Follow-up observation on the goal migration added here: this is not a recurrence of the original “goal was not copied” bug, but there still appears to be a race/crash-consistency gap in the migration implementation.

The migration reads the parent metadata, writes the child, and clears the parent as separate operations. Real SQLite-backed concurrency and failure tests reproduced:

• a child appearing after the absence check being overwritten;
• a child write surviving a failed parent clear, leaving both the parent and child goal records active.

The previously suspected stale-parent-clear case did not reproduce: clear_goal() reloads the parent immediately before marking it cleared, so a concurrent parent update remained unchanged while the child received the earlier snapshot.

A narrow hardening would be an atomic compare_and_move_meta operation that:

• compares the parent against the exact expected value;
• requires the child key to be absent;
• writes the child and removes the matching parent in one transaction;
• makes no changes if either precondition fails.

Focused regressions would cover a child appearing after the absence check, an interruption after the child write, and successful movement. This would preserve the behavior landed here while tightening only the remaining race boundary.