fix(compaction): plugin engines never received the session store, making skew persistence inert

[Kyzcreig]

The calibration stack was correct and completely inert

Three PRs shipped it -- #529 (survive a restart), #539 (key by provider/model), #541 (per content class). All three are right. None of them did anything, because the engine this fleet runs never received the session store.

Measured on the live tree 2026-08-09:

COMPACTION_SKEW ... ratio=1.405 ... class=tool     <- loop running fine
COMPACTION_SKEW ... ratio=1.353 ... class=tool
COMPACTION_SKEW ... ratio=1.335 ... class=tool

sqlite> SELECT COUNT(*) FROM compression_skew_calibration;
0                                                  <- nothing persisted

Root cause

agent_init binds the session store through:

getattr(agent.context_compressor, "bind_session_state", None)

ContextCompressor defines that method. The ContextEngine ABC did not. So every plugin engine skipped the bind silently -- no exception, no log -- left _session_db unset, and made _persist_skew_history() return at its first guard forever.

The consequence is the exact gap #529 was built to close: the learned ratio (a measured 1.33-1.41 under-count on tool output) was discarded on every restart and relearned from raw-rough 1.0.

Fix

Define a minimal default bind_session_state on the ABC.

Deliberately not a fix to LCM alone -- any future engine now inherits a working bind instead of failing the same silent way. ContextCompressor keeps its richer override (cooldown + failure-streak rehydration), pinned by a test so the default can't shadow it.

Proof

E2E on the real engine, not a double:

load_context_engine("lcm").bind_session_state(db, "s1")
_persist_skew_history()
-> [('claude-apr', 'claude-opus-5', [1.4, 1.35, 1.33])]

9 tests, RED-proven -- removing the method fails 7. Two are class-level guards rather than instance tests:

• test_every_concrete_engine_can_be_bound enumerates ContextEngine subclasses, so the next engine to miss this fails at test time instead of in production
• test_agent_init_binds_through_getattr pins the call shape the fix depends on, so the ABC default can't quietly become decorative

The tests assert the wiring, not just the logic. A persist method that works in isolation while nothing hands it a DB is exactly the failure that shipped three times in this subsystem.

979 passed across the compaction / skew / context-engine surface.