[pengyuzhang]

What

Thirteen MCP rows name a declaration site and a launch form but carry no path of their own. manifest.load() now fills that path in from the row that does declare one.

Why

The fourteen M-SITE-* rows enumerate the declaration sites and where each lives per OS. The M-PIN-* rows reuse those sites by name, and M-SP-02 deliberately spans two of them via sites rather than site.

Left unresolved, those rows have no location at all — so a runner cannot write them and a scorer cannot match them, for reasons that have nothing to do with the collector under test. On Linux that is the difference between 51 and 64 executable entries.

How

_resolve_sites() builds a site → per-OS path map from the rows that declare one, then fills it into rows that name a site and lack a path. It runs before validate(), so invariant checks see resolved rows.

Two deliberate choices:

• In the loader, not in each consumer. The runner, the scorer and the synthesizer all need the same answer; resolving once means they cannot disagree about where a declaration lives.
• Fills gaps only. A row stating its own path keeps it — M-SITE-01 still resolves to ~/.claude.json.

Tests

Five added to test_manifest.py, covering the gap this change opened: every declaration resolves a path somewhere, a pinned row inherits its site's path, a row with its own path keeps it, a multi-site row resolves, and resolution stays per-platform.

$ python3 -m unittest tests.test_manifest
Ran 14 tests in 0.011s
OK

Additive only — no existing line changed, and no dependency beyond the standard library.