feat(memory): ACON category guidelines, Memex tool archive, RL admission control

[bug-ops]

Summary

Implements three P2 research-driven memory improvements:

• Memex (research(memory): Memex(RL) indexed experience memory — compact trajectory + external KV store for long-horizon agents (arXiv:2603.04257) #2432): archive tool output bodies to SQLite before compaction; inject UUID references as postfix after LLM summarization so they survive; opt-in via [memory.compression] archive_tool_outputs = false
• ACON (research(memory): ACON agent context compression — 26-54% peak token reduction via guideline optimization (arXiv:2510.00615) #2433): per-category compression guidelines (tool_output, assistant_reasoning, user_context); lazy per-category guideline updates to avoid 3x LLM overhead; opt-in via categorized_guidelines = true
• RL admission (research(memory): RL-based admission control — MemAgent learned write-gate replaces static threshold (arXiv:2603.04549) #2416): AdmissionStrategy enum (heuristic | rl); records all messages (admitted + rejected) to eliminate survivorship bias; pure-Rust logistic regression (50-epoch mini-batch GD); was_recalled wired from SemanticMemory::recall() for positive labels; startup warn when rl is configured (data layer complete; live prediction wired in follow-up)

Architecture notes

• C1 (placeholders destroyed by LLM): fixed — archived refs injected as postfix after summarization
• C2 (archive vs overflow lifecycle): fixed — archive_type column on tool_overflow; cleanup skips archives
• C3 (survivorship bias): fixed — all messages recorded before admission decision

Migrations

• 054_memex_and_acon_categories.sql — archive_type on tool_overflow, category on compression_failure_pairs / compression_guidelines, updated UNIQUE constraints
• 055_admission_training_data.sql — admission_training_data, admission_rl_weights tables

Test plan

• 7302 tests pass (cargo nextest run --workspace --features full --lib --bins)
• cargo +nightly fmt --check passes
• cargo clippy --all-targets --features full --workspace -- -D warnings passes
• 8 unit tests for classify_failure_category() (tool_output, assistant_reasoning, user_context, priority ordering)
• 4 async integration tests for category-aware store methods
• Logistic regression unit tests (gradient descent, prediction, convergence)
• Memex archive/cleanup tests

Closes #2433
Closes #2432
Closes #2416