fix: exclude superseded memories from recall and classify all markdown as docs

[ch405canova-sudo]

Summary

Two real bugs found while operating the server against a live memory store:

1. Superseded entries still surfaced in semantic recall

internal/memory/read.go — the Recall loop never filtered entries whose superseded_by column is set (e.g. after a merge via MarkOutdated/merge flow). A merged duplicate kept ranking as a top result.

Fix: skip entries with m.SupersededBy != "" in the recall loop. They remain visible to List/ListLightweight so maintenance tools can still see the temporal history.

2. Plain markdown files silently dropped from the RAG index

internal/rag/documents.go — classifySourceType was called with empty content (classifySourceType(relPath, "", "")), so the .md heuristic relying on strings.Contains(contentLower, "# ") never matched. A normal .md file (not named readme.md, not under a docs/ path) classified as "" → 0 chunks, silently ignored.

Fix: every .md file classifies as "docs". Added regression cases to TestClassifySourceType.

Verification

• go test ./internal/rag/ passes (incl. new cases)
• End-to-end: a non-README .md file now produces a chunk in the vector store; recall no longer returns the superseded duplicate

Files changed

• internal/memory/read.go
• internal/rag/documents.go
• internal/rag/rag_test.go
• .gitignore

Fixes #18
Fixes #19

[ch405canova-sudo]

This work is a joint effort by chaos (@ch405canova-sudo) and opencode — found and fixed together on a live llama.cpp + agent-memory-mcp stack (August 2026).

[ipiton]

Thank you for this — both reports were accurate, and the root-cause analysis in #18 and #19 pointed straight at the right lines.

Status of each half:

Markdown classification (#19) — already fixed on main in b052f2d / bd00d87, before this PR arrived; the .md branch now returns docs unconditionally. That is what makes this PR conflict: it changes the same lines. The fix has not shipped in a release yet, which is why v0.9.1 still reproduces it. Your two regression cases are worth keeping and have been picked up.

Recall filter (#18) — a real bug, still present on main, and your fix is the right shape. It landed in #20 with one correction: the successor is now looked up in the cache rather than trusted, because Delete does not clear superseded_by on predecessors. With an unconditional continue, deleting a successor buries its predecessor permanently — an archived entry beats no entry at all. #20 also updates TestMarkOutdatedDownranksMemory, which pinned the opposite contract and would have failed CI here, and adds coverage for the recall path itself (this PR only touched the RAG tests).

#20 supersedes both halves, so there is nothing left to rebase here — feel free to close this PR. The credit is yours, and the report quality is genuinely appreciated. Happy to take further PRs.