v2.31.0 — memory-system upgrades (importance scoring + lessons-synthesis)

Implements the two high-value memory-system upgrades from the F18 audit finding ("3-stage agent memory"). Deeper machinery (vector search, spreading activation, hierarchical memory, a formal procedural tier) was deliberately left out as over-engineering for this filesystem-backed, human-in-the-loop design.

Added

• Importance scoring on plans/LESSONS.md — optional inline [I:N] tag (1-5; omitted = 3). The CLOSE archivist now trims the 200-line cap by importance-then-recency and never drops [I:5] lessons.
• lessons-synthesis emit-template slug (16 → 17 templates) — a structured CLOSE-time reflection guide (Recurring Patterns / Failed Approaches / Successful Strategies / Codebase Gotchas, each [I:N]-tagged) the archivist uses to promote recurring per-plan findings/decisions into LESSONS.md.

Fixed

• Corrected SKILL.md File-Lifecycle/CLOSE wording that conflated LESSONS.md trimming with SYSTEM.md's demote-by-staleness rule.

Test suite unchanged at 295/295 (the new slug is auto-covered by existing loop tests — no new test blocks). make validate + make test green.