[G1] Hydration reports capability "full" unconditionally for local sources; compute it from covered evidence kinds #169

Part of #160 (group 1 — independent, small). This is a correctness bug in the current release, independent of the migration, and it blocks every parity issue from being verifiable through the public API.

Problem

build_result in crates/ai-hist/src/hydrate.rs:1642-1650 hard-codes:

capability: "full".to_string(),
discovery_state: "full".to_string(),
presence: "local".to_string(),

So a Cursor, Grok, or OpenCode session that produced prompts only (no session_events, tool_calls, file_edits, usage) is reported to every consumer — CLI, TS SDK, MCP hydrate_session, and the SDK's merge ranking { full: 2, partial: 1, shallow_only: 0 } (sdk-ts/src/operations.ts:945) — as full. This is exactly the "well-formed answer computed over nothing" failure class from CLAUDE.md.

The correct rule already exists for the plugin path: source_intake::apply_normalized (crates/ai-hist/src/source_intake.rs:~270) sets full only when every kind in FULL_SESSION_KINDS (crates/ai-hist-core/src/source_evidence.rs:19: History, SessionEvent, ToolCall, FileEdit, Relationship) is in covered_kinds; otherwise partial. The local path never applies it.

Scope

1. Define per-source declared coverage in the provider registry (crates/ai-hist/src/sources.rs / discover.rs ShallowSessionProvider): a fn evidence_kinds(&self) -> &'static [EvidenceKind] that says which kinds the local parser can produce for that source (today: claude/codex = all five; cursor/grok/opencode = History only; relay = none). This is a property of the provider, not of a database, matching how relationship_capabilities is exposed.
2. build_result computes capability = full if declared coverage ⊇ FULL_SESSION_KINDS, else partial; and discovery_state = the value actually written to the sessions row (hydrate.rs currently upgrades it separately — reuse that value instead of a literal).
3. Add a coverage: Vec<EvidenceKind> field to HydrateSessionResult (and HydrateSessionResult.diagnostics gains HYDRATION_PARTIAL_COVERAGE naming the missing kinds), so a consumer can tell which evidence is absent. Bump SESSION_HYDRATION_CONTRACT_VERSION (hydrate.rs:11) → 3; update sdk-ts/src/contracts.ts:162-170, normalization.ts normalizeHydration, the MCP tool description, and docs/session-catalog.md.
4. sessions.discovery_state semantics: a partial hydration must not set discovery_state = 'full' in a way that makes the unchanged short-circuit (hydrate.rs:187-201) skip re-hydration after a later parser upgrade — verify the parser_version check still forces a re-parse.
5. Each parity issue (Cursor, Grok, OpenCode) flips its provider's declared coverage as part of its own PR.

Acceptance

• hydrateSession({source:'cursor'}) on the current parser returns capability: 'partial', coverage: ['history'], and a HYDRATION_PARTIAL_COVERAGE diagnostic listing session_event, tool_call, file_edit, relationship.
• hydrateSession({source:'claude'}) still returns full.
• sdk-ts/src/session-evidence.test.ts / local-history-contract.test.ts cover the new field and version.
• CHANGELOG.md entry under Breaking: "hydration contract 3; local hydration no longer claims full for prompt-only providers".

Out of scope

Making the prompt-only providers complete (their own issues).