feat(KR-7b): swap C2 capability mirror for kora__read_kora_capability_row MCP fetch — closes D-kr2-st2

[rafe-walker]

Summary

KR-7b swaps the hand-mirrored C2 capability dict for an authoritative MCP fetch via kora__read_kora_capability_row at provider initialize. Closes D-kr2-st2-capability-matrix-mirror.

CC#3 elected Option A (use the KR-7a-wired IsoKronMCPClient) over Option B (parallel httpx fetch path) since KR-7a is shipped — one canonical MCP-call pattern across all closure swaps. The hand-mirrored 49-entry dict stays as a dev/test fallback (parity test still guards drift); production fetches authoritatively at start.

Production-test posture (same as KR-7)

K-7's kora__read_kora_capability_row handler is a notImplementedHandler stub on substrate main; substrate-team's dispatch tier (queued, task NousResearch#395) un-stubs the handler. KR-7b's code shape is sound; mock tests verify the populate machinery; production deploys wait on the dispatch tier landing. On fetch failure the hand-mirrored fallback stays in place + a [kora.capability_matrix.fallback] WARNING is logged — dev/test ergonomics survive substrate downtime; sessions still run. Operators grep [kora.capability_matrix.fallback] in production logs to confirm fetch health.

What landed

File	Change
plugins/memory/isokron/capability_matrix_mirror.py	New populate_capability_matrix_from_mcp(mcp_client) coroutine — invokes kora__read_kora_capability_row (no args), validates {capability_matrix: dict[str, bool]} shape (defensive checks on both keys and values), mutates ACTOR_CAPABILITY_MATRIX_KORA_COLUMN in-place via clear() + update(). Returns the entry count. New CAPABILITY_MATRIX_MCP_TOOL constant. Module docstring rewritten: hand-mirrored data is now the dev/test fallback (no longer "C2 interim").
plugins/memory/isokron/provider.py	initialize() now calls new _refresh_capability_matrix_from_mcp() after starting the connection. Two log paths: INFO [kora.capability_matrix] on success; WARNING [kora.capability_matrix.fallback] on MCP-fetch or transport failure (fallback intact).
tests/plugins/memory/test_capability_matrix_mcp_fetch.py	new — 11 tests covering populate (happy + 6 defensive error paths) + provider.initialize (INFO success, two WARNING fallback variants). Autouse fixture snapshots + restores the module-level dict for xdist isolation.
tests/plugins/memory/test_provider_end_to_end.py	_FakeMcpClient.invoke routes by tool name — returns a canonical 3-entry capability_matrix for kora__read_kora_capability_row; mock event_id for kora__append_event. E2E asserts the initialize-time fetch fired and replaced the dict. New restore_capability_matrix fixture.
BUILD_DEVIATIONS.md	D-kr2-st2 moved Open → Closed with KR-7b's Option-A rationale, production-test posture, hand-mirrored fallback retention note.
plugins/memory/isokron/README.md	Deferred-surface table trimmed 3 → 2 rows. D-kr2-st2 moves to Recently-closed. Individual pitfall rewritten as "MCP-fetched at provider initialize" with fallback semantics.

Why hand-mirrored fallback stays

The 49-entry C2 dict at module import is now framed as the dev/test fallback rather than "C2 interim":

• Production: MCP fetch succeeds at initialize() → dict replaced with authoritative substrate data. K-13's upcoming additions (cap_emit_chain_event, cap_write_relationlink) flow through automatically at next provider start — no Python-side code change.
• Dev/test: tests run without a live substrate. Fallback provides 49-entry coverage so actor_has_capability lookups still resolve. K-13 will still require a 2-line bump to the fallback for dev parity, but production picks it up automatically (the parity test against TS source catches the bump-needed signal at CI).
• Production with substrate downtime: fallback stays + [kora.capability_matrix.fallback] WARNING fires. Sessions keep running on (potentially stale) dev data rather than failing closed. Operators grep the tag to confirm fetch health.

Test plan

11 new tests in test_capability_matrix_mcp_fetch.py, all passing:

populate_capability_matrix_from_mcp (7):

• Invokes kora__read_kora_capability_row with empty args
• Replaces the module dict in-place; identity preserved (forward-stability for capability_check consumers)
• Raises ValueError on None mcp_client
• Raises RuntimeError on missing capability_matrix key
• Raises RuntimeError on non-dict response
• Raises RuntimeError on non-bool capability value (defensive vs substrate regression)
• Raises RuntimeError on non-str capability key
• Propagates underlying MCP-side exceptions unchanged

provider.initialize (3):

• Success: INFO [kora.capability_matrix] populated from kora__read_kora_capability_row — N entries + dict replaced
• Substrate-down: WARNING [kora.capability_matrix.fallback] MCP fetch ... failed (...); using hand-mirrored fallback + fallback intact
• Transport-unavailable: WARNING [kora.capability_matrix.fallback] could not reach MCP client (...)

E2E updated (test_provider_end_to_end.py): _FakeMcpClient.invoke routes by tool name. Asserts the initialize-time fetch fired with the spec-pinned tool name + empty args, dict was replaced with the canonical test matrix, and the 2 append_event emits still fired separately.

Gates

• ty check — 7,337 diagnostics, zero-delta vs KR-7 baseline. All targeted-file diagnostics resolved.
• pytest tests/plugins/memory/ — 362/362 passing (11 new KR-7b + 351 pre-KR-7b).
• Full suite via xdist (-n auto): 24,652 passed / 175 failed / 129 skipped. Δ vs KR-7 (24,561 / 203 / 52err / 129): +91 passed, −28 failed, −52 errors. Same tests/tools/* + tests/tui_gateway/* xdist isolation noise; none touch plugins/memory/isokron/.

Rule-6 / BUILD_DEVIATIONS / Open asks

Closed: D-kr2-st2-capability-matrix-mirror (KR-7b).

2 deferrals remain open — both unblocked at substrate; PM has dispatched the matching closure swaps:

Deviation	Closes when
D-kr2-st3-no-scratchpad-write-mcp-tool	KR-8 (K-8 merged bd165eb2; PM dispatched)
D-kr3-st2-no-relationlink-write-mcp-tool	KR-9 (K-10 + 0083 in prod; PM dispatched)

Both swaps follow the KR-7 template — same mechanical pattern, same production-test posture. Shipping KR-8 next per PM-suggested order, then KR-9. After all 4 closures merge, the standing-by state becomes dispatch-tier-gated (substrate-team task NousResearch#395 + service-token mint for production).

🤖 Generated with Claude Code