modelInfo contract follow-up: grant coherence + §10.1 model-field optionality needs a spec decision

[antra-tess]

Deferred from PR #79 by antra's call ("let's not nitpick modelinfo and leave it as an open issue"). Sol's final review point, recorded verbatim so nothing is lost:

1. Grant coherence: AF no longer advertises modelInfo (no truthful contextWindow source exists), but computeGrant can still grant modelInfo from a server's advertisement — producing an effective grant that promises a method the host then answers with -32601. The behavior is fail-safe (explicit error, never a hang or a fabricated partial) but incoherent: a granted path should be an answerable one. Fix shape: hosts exclude capabilities they do not implement from the grant computation (a host-implementation mask alongside the config mask), or the grant carries them and §12 documents that a granted modelInfo may still answer method-not-found when the host cannot source truthful values.

2. §10.1 model-field optionality is a SPEC question, not a code question: PR MCPL 0.5: capability grants, negotiated policy, enforcement (#76) #79 made McplModelInfo.vendor/contextWindow/capabilities optional and sends {id} in beforeInference. SPEC §10.1 currently implies the full shape; the honest-over-fabricated principle (§12 discussion in PR MCPL 0.5: capability grants, negotiated policy, enforcement (#76) #79) supports optionality for hook params specifically, but that needs a sentence in SPEC §10.1 rather than a silent host divergence — the same examples-are-load-bearing lesson as RFC-001's featureSets array.

Both halves are small; neither blocks the 0.5 release train per antra's ruling. Review context: #79 (comment)