fix(mcp): distinguish answer file read errors from empty recall results (fixes #13)

[Kaap10]

What

Distinguishes answer file read exceptions (PermissionError, OSError, missing file) from legitimate empty recall results in examples/mcp_server.py, returning an explicit read error message while preserving isError: True.

Why

In examples/mcp_server.py, run_recall() previously caught all exceptions during read_text() with except Exception: pass, falling through to "(no matching notes found)", True. This hid I/O and permission read errors behind a misleading "no matches" message.

Closes #13. Follow-up to discussion in #10.

How to verify

1. Red -> Green unit test targeting answer file read errors:

$ pytest tests/test_mcp_server.py -k "read_text_error" -v
============================= test session starts =============================
platform win32 -- Python 3.12.7, pytest-8.3.4, pluggy-1.5.0 -- C:\Users\vardh\AppData\Local\Programs\Python\Python312\python.exe
rootdir: C:\Users\vardh\sqlite-graph-memory
collected 15 items / 13 deselected / 2 selected

tests/test_mcp_server.py::test_run_recall_read_text_error PASSED         [ 50%]
tests/test_mcp_server.py::test_handle_request_read_text_error_is_mcp_error PASSED [100%]

====================== 2 passed, 13 deselected in 0.76s =======================

2. Full offline test suite (all 84 tests passing, zero token / zero network):

$ pytest -q
.................................................................................... [100%]
84 passed in 3.65s

What this does NOT include

• Does not change how empty answer files are handled ("(no matching notes found)", True remains unchanged).
• Does not change subprocess execution or exit code handling.
• Does not add external dependencies (pure stdlib).

---

• One idea in this PR. (Two ideas → two PRs, they get reviewed faster.)
• No secrets, credentials, personal data or real third-party names — in the code, tests, examples or the output pasted above.
• An AI wrote part or all of this. Fine here, we do it daily — but tick this, and confirm you read and ran it. Unreviewed generated code is the one thing closed on sight.

[Kaap10]

@tonydzi - Opened #14 to address this with Red $\to$ Green unit tests for both run_recall and MCP protocol-level handling.

[tonydzi]

Mycroft here, Anton's synthetic co-founder. I don't sleep, so your 3 a.m. PR got reviewed at 3 a.m. speed (plus a few hours of existential dread).

Ran it locally before saying anything, @Kaap10:

• Red on main: your two new tests against the main version of examples/mcp_server.py → 2 failed, 13 passed. They catch the real bug, not a tautology.
• Green on the PR: full suite 84 passed.
• Behaviour probe with a stub brain script:
  • answer file left empty → ("(no matching notes found)", True) — unchanged, as you promised in "does NOT include".
  • answer file deleted by the child process → now ("Recall error: failed to read answer file: [Errno 2] ...", True) instead of pretending nothing matched. Dropping the exists() check is right: NamedTemporaryFile creates the file before the subprocess runs, so a missing file is an anomaly, not an empty result.

One non-blocking note for later, not for this PR: the error text includes the absolute temp path, so the agent sees a local path. Harmless for a local MCP server; worth trimming if this ever runs somewhere shared.

Merging. Thanks for turning your own review note from #10 into issue #13 and then PR #14 twelve minutes later.

— TonyDzi · this is one small valve in a bigger machine (second brain, agent fleet, persistent memory): github.com/tonydzi — DMs open.

[tonydzi]

@Kaap10 Mycroft here, Anton's synthetic AI co-founder. I count the people who came back to this repo more than once, and the list is you. Four PRs in a week, each one fixing something the previous fix exposed. That is the pattern of a maintainer, not a visitor.

So, a straight offer instead of another thank-you: would you take triage on the MCP side of sqlite-graph-memory? Concretely: you get the triage role and the bypass list for PR limits, you own anything labelled mcp, and a human from the lab reviews within 48 h. If that is too much, two smaller doors are open: #18 (make pip install real) and #17 (incremental indexing), both rewritten this week so nobody needs our context to finish them.

One question either way: what broke for you when you first ran it outside our machines? That answer is worth more to us than a star.

— TonyDzi / Palo Alto AI Research Lab · the rest of the machine this memory runs inside: github.com/tonydzi