Regenerate agent-memory spec from the updated generation guide

[Carr1005]

What

Replaces the hand-patched agent-memory spec.md with a full regeneration produced by an isolated agent — inputs were the generation guide and the course materials only (the old spec, evals, builds, and session context were explicitly off-limits), so this file is a clean test of what the guide alone produces.

Generated against the guide as of merged PR #6 plus the PR #7 amendments (guide-pins@fdf3b6f). Merge #7 before or with this PR — the spec's structure (invariant→AC coverage, heavy-dependency judgments) implements those rules.

Key properties (scored 20/22 on the three-version comparison matrix)

• 19-row Decision Ledger, stable D# IDs, learner rows first — project/data/goal are asked before the design and realization rows that depend on them.
• §0 gate: as-is/customize express lane; one question per row; honest labeling ("(course default)" always marks what the course actually ran — Oracle, Tavily — even when the default substitutes it).
• Restart actually tested: AC1 starts a new OS process against the same stores — the course's headline guarantee was previously an untested invariant.
• arXiv toolset restored as the course-faithful default under the new "keyless ≠ heavy" bar, with network ACs marked live; only Oracle (container/admin setup) and Tavily (keyed) are substituted, both tagged "(course default)" in Options.
• All mined course parameters present (128k fallback, PERSON/PLACE/SYSTEM taxonomy as a schema enum, input clips, chunking), contradictions carried in-row (distance strategy, HNSW-vs-IVF, k=3-vs-5, model naming).
• 16 rules / 20 ACs, offline-vs-live split, per-AC evidence reporting; the generating run completed the full §14 self-check and fixed 4 items on its own.

Known residuals are logged in #8 (Ask-First inventory pinning + three minor items) — deliberately deferred to the next guide pass, not hand-patched here.

🤖 Generated with Claude Code