[BUG] Silent bulk handle eviction reports "not found" — user state vanishes with no warning #532

Severity: P1 — silent loss of user state. Verified live 2026-08-09 (server at main cbac09f).

Problem

Both handle stores silently bulk-evict the oldest entries when they hit their cap, and the caller only learns about it when a later call returns a generic "not found" — indistinguishable from a typo.

Live repro: a data handle created at the start of a session was gone after ~52 subsequent load_data_source calls:

inspect_data(data_handle="data_bdaec9a2")
  -> {"success": false, "error": "Data handle 'data_bdaec9a2' not found"}

Root cause

• Estimator store — HandleManager._cleanup_oldest (handles.py) deletes the 10 oldest outright when len >= max_handles (default 100).
• Data store — Executor._cleanup_oldest_data (executor.py) deletes the oldest ~20% at once, emitting only a logger.debug — nothing reaches the caller.

Proposed fix

Distinguish evicted from never existed. Keep a tombstone set in both stores and return, e.g., "Data handle 'X' was evicted (limit N reached); reload the source." Better still: expose the limits + current usage via list_handles / list_available_data, and/or make eviction opt-in (raise at the cap). Never silently discard user state.

Re-verify

Create handles past the cap, then use the first one — the error must say evicted, not not found.

(Catalogued as BUG-08 in MCP_TEST_FINDINGS.md.)