SDK-tier scoring (SEP-2484)

The 2026-07-28 spec ties the publicly visible SDK tier to the conformance suite's MUST / SHOULD pass counts. mcptest aggregates a run's results into a single tier so a CI badge can summarize the run with one letter and a percentage.

Tiers

Tier	Badge	Rule
Tier 1 (gold)	`T1`	Every MUST check passes and at least 95 % of SHOULD checks pass.
Tier 2 (silver)	`T2`	Every MUST check passes and at least 70 % of SHOULD checks pass.
Tier 3 (bronze)	`T3`	Every MUST check passes. SHOULD coverage below 70 % but no MUST regressions.
Fail	`F`	At least one MUST check failed. No tier awarded.

The thresholds match the spec table exactly. MAY checks affect neither the tier nor the percentage; they are reported for the operator and never gate the badge.

Library surface

mcptest_core::conformance::tier:

Item	Purpose
`TierInput { must_passed, must_total, should_passed, should_total }`	Aggregated counts from a run.
`Tier { Tier1, Tier2, Tier3, Fail }`	Verdict.
`score_tier(input) -> Tier`	Pure function implementing the SEP-2484 rule.
`Tier::badge()`	Short letter form (`T1` / `T2` / `T3` / `F`) for CI badges.
`TIER1_SHOULD_THRESHOLD`, `TIER2_SHOULD_THRESHOLD`	The two pass-rate boundaries as `f64` constants.

Edge cases

A corpus with zero MUST checks lifts the run to Tier 1 (there is nothing to fail, so the floor is the highest passing tier).
A corpus with zero SHOULD checks treats the SHOULD pass rate as 1.0 (vacuously satisfied), so the run lifts to Tier 1 when MUST is clean.

Vendored corpus

The MCP conformance suite scenarios live under conformance-corpus/ in this repo, not as a submodule. The README in that directory documents the refresh procedure: locate the upstream working-group repository, copy the scenarios for the target spec revision, update the upstream tag in upstream.txt, and open a pull request that lists added, removed, and changed scenarios so a reviewer can decide whether the change crosses an SDK-tier boundary.

A weekly cron job opens the refresh PR automatically once the upstream repository is identified. Until then, the directory is intentionally empty.

Planned follow-up

mcptest conformance run and mcptest conformance refresh subcommands. The full contract (flags, corpus resolution order, fetch transport, exit codes) lives in conformance-cli.md.
The refresh GitHub Action that opens a weekly PR against conformance-corpus/.
A documented mapping table between mcptest's conformance packs (mcptest-core::compliance rules) and the upstream SEP check ids so a failing requirement points the operator at the right pack.