fix(serve): reorder except clauses so the corrupted-graph message is reachable

[kimdzhekhon]

Summary

• json.JSONDecodeError subclasses ValueError, so _load_graph()'s except (ValueError, FileNotFoundError) clause always matched first — the except json.JSONDecodeError clause with the "graph.json is corrupted (...). Re-run /graphify to rebuild." recovery message was unreachable dead code.
• Users hitting a truncated/corrupted graph.json got the bare underlying json.JSONDecodeError message instead, contradicting what SECURITY.md documents for this exact threat.
• Moved the json.JSONDecodeError clause first so it actually catches. Audited the rest of serve.py's exception handlers for the same narrower-after-broader ordering bug — found no other instance.

Fixes #2005.

Test plan

• Added test_load_graph_corrupted_json_prints_recovery_message — writes invalid JSON, asserts the recovery message (not the bare decode error) is printed
• uv run pytest tests/test_serve.py -q -k load_graph — 6 passed
• uv run pytest tests/ -q — full suite: 3412 passed (5 pre-existing failures unrelated to this change, missing optional openai dependency in this environment)

[Job28703]

Independently reproduced the fail-then-pass: checked out v8 base (unfixed) with the new test applied — test_load_graph_corrupted_json_prints_recovery_message fails exactly as expected (bare JSONDecodeError message instead of the recovery hint), confirming json.JSONDecodeError is indeed swallowed by the broader (ValueError, FileNotFoundError) clause ahead of it. With the reorder applied, the same test passes (6 passed on test_serve.py -k load_graph). Also spot-checked the rest of serve.py's ~20 except clauses — no other narrower-after-broader ordering issue. Clean, well-scoped fix. Thanks!

[HerenderKumar]

Came here from #2005 — I had claimed it on the issue a few hours before this PR went up and had the same fix built locally, so I can vouch for the approach: json.JSONDecodeError subclasses ValueError, so moving its clause first is exactly right. I ran the same change against the full serve suites locally (117 passed, 1 skipped).

One suggestion: a small guard test asserting that a non-decode ValueError (e.g. a non-.json path) still prints the generic error: ... message rather than the corrupted-graph hint — it pins the clause order so a future refactor can't "simplify" it back:

def test_load_graph_generic_value_error_message_unchanged(tmp_path, capsys):
    p = tmp_path / "graph.txt"
    p.write_text("not a graph")
    with pytest.raises(SystemExit):
        _load_graph(str(p))
    err = capsys.readouterr().err
    assert "must be a .json file" in err
    assert "corrupted" not in err

LGTM either way.

[kimdzhekhon]

Good catch — added that exact guard test. Thanks for vouching for the approach and for validating against the full serve suite.