feat(conformance): wire make testconf-events against the fork's Events suite

[panyam]

What changes

Adds a make testconf-events target that drives examples/events/kitchen-sink against the two phase-1 Events scenarios proposed upstream in modelcontextprotocol/conformance#504, plus the local-suites.yaml manifest entry and the regenerated path defaults. The suite runs red on purpose: 27 pass / 13 fail / 5 warn, and every failure is a divergence between our events implementation and the design sketch that merged upstream on 2026-09-08.

Wired as INFO at stage 8i, following testconf-skills.

Prerequisite knowledge

• conformance/NOTES.md — how the testconf-* targets are wired and which upstream changes to watch. The MCPCONFORMANCE_*_PATH convention this target follows is described there.
• Root CLAUDE.md § Cross-cutting rules — the path-defaults.{mk,sh,just} generation rule. Hand-editing two of the three fails CI as case E drift, which is why gen_conf_paths.py runs here.
• Events design sketch — what the suite scores against. Skim the Summary and Cursor Lifecycle sections to read the failures.
• experimental/ext/events/NOTES.md — our side of the same surface, including the earlier spec-alignment sweeps.

Reviewer's guide

A restaurant can cook a dish that is not on the menu. The kitchen has the ingredients and the line knows the recipe, so if you somehow know to ask you get it, but nobody reading the menu ever orders it. Our events server is that dish. events/list answers, all three delivery modes work, and capabilities.events appears nowhere, which is the menu a spec-following client reads before deciding what to ask for. That is the single most consequential thing this target surfaces, and it is the kind of gap that stays invisible precisely because every test we wrote ourselves already knows to ask. The rest of the failures have the same flavour: places where our implementation is self-consistent and internally tested, and diverges from the document anyway. Three of the six were unknown before this run. So read the diff as plumbing, and spend the attention on whether INFO is the right status for a target whose whole value is being red.

Read in this order:

1. conformance/local-suites.yaml — start here. The note is the finding, recorded with its date and measurement method.
2. conformance/scripts/conf-events.sh — the runner. Spawns the fixture on :18101, runs the two scenarios, prints each summary, exits 0 and explains why.
3. conformance/Makefile, Makefile, justfile, conformance/justfile — target plus the three mirrors. just is the one people forget.
4. scripts/testall.sh — the script map entry and run_stage 8i ... info.
5. conformance/path-defaults.{mk,sh,just} — generated, skip.

Decision log

• INFO, not a gate. Two reasons, both outlasting this PR. The suite scores against a design sketch with no SEP number, so its check ids still carry a placeholder sep-9999- prefix that will be renamed. And the failures are ours to fix, so gating would block the umbrella on work this PR deliberately does not do. The precedent is testconf-scope-challenge, INFO for tracking an unmerged PR head. Flip to a gate once the divergences close and the spec text stabilises.
• Points at the fork branch while the upstream PR is in review. feat/events-conformance-suite on panyam/mcpconformance, default path ../conf-events, matching how every other fork-based suite here resolves. The scenarios are proposed upstream as a draft (feat(sep-xxxx): Conformance suite for triggers and events modelcontextprotocol/conformance#504); retarget source.branch to upstream main once it lands, the same way testconf-scope-challenge is meant to.
• Fixture defaults left alone. The synthetic feeders emit every 2-4s, which is what gives the EventOccurrence checks a chance to grade something real. A silenced fixture reports them untestable, which is correct but tells us less.
• No implementation fixes in this PR. The six divergences want their own changes, some wire-breaking. Bundling them would bury the wiring and make the target's first run green for the wrong reason.

Risk / blast radius

• Affects: make testall gains stage 8i. Because it is run_stage ... info the stage cannot fail the umbrella, and conf-events.sh exits 0 unconditionally.
• Needs a sibling worktree. ../conf-events on feat/events-conformance-suite, or MCPCONFORMANCE_EVENTS_PATH set. Absent, require_conf_dir fail-fasts with a remediation message like the other fork-based suites.
• Tests covering this: uv run scripts/check_local_suites.py reports no drift across 9 suites, which is the gate that would catch a missing mirror or a wrong stage label.
• Be paranoid about: the four mirrors. A target added to conformance/Makefile and not to justfile passes local make and fails the drift check. All four were touched here; confirm.
• Related: rebased onto main after chore(conformance): retarget testconf-skills to upstream main, gate it, and move the renderer to pnpm #1377, feat(events)!: align wire fields with the merged design sketch (nextPollMs, maxAgeMs, inputSchema + validation) #1379 and feat(events): emit list_changed and terminate subscriptions on event-type removal (28ec35e9) #1381. chore(conformance): retarget testconf-skills to upstream main, gate it, and move the renderer to pnpm #1377 promoted testconf-skills from info to a gate, which conflicted with this branch's testall.sh line; resolved in favour of main, so skills stays gated and events is added as info alongside it. feat(events)!: align wire fields with the merged design sketch (nextPollMs, maxAgeMs, inputSchema + validation) #1379 and feat(events): emit list_changed and terminate subscriptions on event-type removal (28ec35e9) #1381 moved the numbers, which are re-measured above. Expect the same again: any events fix restales the counts in local-suites.yaml, so re-run and update the note.

Before / after

Before, nothing scored our events surface against the spec text. After:

$ make testconf-events
Spawning fixture on :18101, scratch dir /tmp/conf-events.XXXX
  events-discovery: Passed: 8/11, 3 failed, 1 warnings
  events-poll: Passed: 19/29, 10 failed, 4 warnings

The six divergences, each observed against the running fixture:

Check	What we do
sep-9999-capability-events-object	answer events/list, declare no capabilities.events
sep-9999-poll-events-array	omit events rather than return an empty array
sep-9999-descriptor-delivery-subset	events.topology sends delivery: null where an array is required
sep-9999-descriptor-input-schema	events.topology carries no inputSchema
sep-9999-poll-mode-unsupported	answer events/poll for events.topology, which advertises no poll delivery

All five are tracked in #1380. The bottom three collapse into one question rather than three fixes: with #1381 having added inputSchema to the three real sources, every remaining descriptor-level failure is the events.topology meta-source not keeping the descriptor contract. Whether a meta-source belongs in events/list as a peer of real event types is worth settling before patching the symptoms.

The first one is the one I would fix first. It is not that a field is wrong; it is that a client reading capabilities never learns the surface exists.

#1379 and #1381 landed while this PR was open and closed three earlier findings: the nextPollSeconds rename (G31), and list_changed plus termination (G34, G30). The suite confirms all three, which is the first useful thing it has done as a regression check rather than a discovery tool.

Out of scope

• Fixing any of the five. All are tracked in events: four wire divergences from the merged design sketch, found by the new conformance suite #1380.
• Phases 2 and 3 of the suite (push, webhook), and the --header runner option needed to score metronome-mcp.fly.dev.

Refs #1374

🤖 Generated with Claude Code

https://claude.ai/code/session_017kyEZdgh1zjdhURgdcZ4nL

[panyam]

Filed the four previously-untracked divergences as #1380. G31 stays with feat/events-spec-wire-alignment, G30 and G34 keep their existing tracking.

Once #1380 lands, the counts recorded in conformance/local-suites.yaml go stale by design — re-run make testconf-events and update the note.