[ViperJuice]

Amends specs/phase-plans-v12.md. The fable seat — the fourth panel seat, filled natively rather than through the CLI — reported after #167 merged. Docs only.

Three roadmap-changing findings, each verified against source before applying.

1. Inline reconciliation self-deadlocks — and that was the fix I had just applied

_index_capabilities awaits _send_request (manager.py:1292). Those futures are resolved by the very read loop that received the notification — pending.future.set_result at :1791 and :2010, the two dispatch functions. Calling reconciliation inline makes the loop await a future only it can resolve. Instant deadlock, not a hypothetical.

#167 named the hazard in EC-FANOUT-6 but didn't mandate a mechanism. It now requires a spawned, per-server-coalesced background task — coalescing also bounds a downstream that spams list_changed.

2. A catalog-freshness criterion was missing

Every notification-shaped criterion in #167 can pass while the gateway still serves the old catalog: get_tool reads self._tools (manager.py:2698), populated only by _index_tools.

New EC-FANOUT-9: after a real downstream emission, a client fetching through the gateway sees the new tool and can invoke it. It is also the only criterion that catches the deadlock above — a deadlocked re-index never refreshes anything, so a slow-subscriber test cannot see it.

3. The wrapper decision is reversed — now deletion

#167 kept is_version_newer as a wrapper and policed it with a test asserting no src/ call site negates it. The seat overturned that on this roadmap's own principle, and the argument is decisive:

My replacement test was itself syntactic. It misses f = is_version_newer; not f(...), x == False, and if is_version_newer(...): ... else: <treat as up to date> — the last of which reproduces the collapse with no negation at all. That is precisely the #163 lesson ("a syntactic check cannot prove a dataflow property") applied to the lint's own replacement. A smaller lint inherits the same unsoundness with fewer lines.

Grep-verified: after EC-TRISTATE-2 migrates refresher.py:235 and :435, both wrappers have zero src/ callers. pmcp's public surface is its MCP tool schemas, not this internal module, so there is no compatibility constituency. is_version_orderable stays — genuinely unary.

Deletion collapses EC-TRISTATE-3 and -4 into absence, and the verification becomes an import-time assertion rather than a grep.

Also

• Assumption 2 now cites its evidence. mcp 2.x opens the standalone GET stream on notifications/initialized (SDK streamable_http.py:566), which pmcp sends at manager.py:2185 — so the remote path is reachable. Also notes fake_remote drives only the remote loop, so stdio needs its own test.
• EC-UPDPATH-4 reframed to "update_server follows the same live-environment contract as every other spawn path", making the uniformity the documented invariant rather than a probe-window special case.

What the seat confirmed

It independently reproduced the other three seats' findings — the index-backed catalog chain, the absence of any per-server dimension in the SDK's SubscriptionFilter (verified: tools_list_changed, prompts_list_changed, resources_list_changed, resource_subscriptions only), and check_staleness as a third identity-blind site.

It also upheld EC-UPDPATH-4's document-don't-freeze decision, with better reasoning than the roadmap had: a frozen update_server would be the only snapshot-semantics spawn path, and a crash-reconnect minutes later would silently un-freeze it — so the guarantee could not be kept even in principle.

Panel tally across both rounds: 4 seats, 4 sets of findings, every one valid. Two of them landed on code I had written in response to an earlier finding.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}