Release the _meta revision fix as 0.2.0

[HackTuah]

Owner decision: the _meta revision fix that merged as 0.1.2 releases as 0.2.0.

Why

For a client declaring 2025-11-25, a method that was refused now answers (ping), and results on that path have lost two fields — resultType and _meta serverInfo. A 0.1.0 consumer could have depended on either. For a published package the wire JSON is the API, so this is a minor bump. The field removal is the breaking-shaped half and carries the decision on its own.

One of the two stated grounds was wrong and is corrected rather than quietly reworded. The rationale said "a method that answered now refuses" — true of no method in this release. A reviewer caught it by enumerating every method on this tree and on 0.1.0's:

ping         0.1.0: REFUSED -32601      0.2.0: answered
every other method: unchanged status on both

That sentence reached the changelog unchecked from the decision that used it. Correcting it strengthens the case for 0.2.0 rather than weakening it.

A standing rule, from a lane finding

The README recommended {:beam_mcp, "~> 0.1"} — a requirement spanning the very break the version bump exists to signal. Nothing was broken for a new user; the defect is that a consumer who copied it at 0.1.0 is carried across the break by a routine mix deps.update with no signal. It survived a deliberate version sweep because the stale string was 0.1, not the 0.1.2 that changed.

So CONVENTIONS.md gains: the README states what the package does today, and every behavioural claim in it is pinned by a test that runs. test/beam_mcp/readme_claims_test.exs discharges it — each test quotes the README sentence it pins and asserts that sentence is still present, so a claim that moves without its test fails, and a test guarding a deleted claim fails too rather than reading as coverage while guarding nothing.

Three claims previously had nothing holding them, including that every method is served bare with tools/call reaching dispatch — a security property that rested on prose.

Verification

./tools/gate.sh    format / compile / test / credo  pass
                   reuse  pass (20 commentable files)
                   licence files  pass
                   Gate OK.  exit=0
mix test           47 tests, 0 failures

Scored by mutation: reverting the README dependency fails the claim quote (8 tests, 1 failure); inserting a session guard fails the two behavioural tests (8 tests, 2 failures). Mutant B's source check reported old remaining: 1 because the replacement embeds the original line — that count carried no information, and the effect is what proved it applied.

Base

PR #7 merged by rebase while round 7 was in flight, so main carries a rewritten commit and the local one was not its ancestor. Trees compared identical (a866ad8d both, empty diff) before resetting onto main.

Review

Rounds 7 and 8, two independent lanes on checkouts of the index; both approved tree 072cf981. The committed tree is that tree plus the two round-8 lane reports and nothing else — verifiable with git diff 072cf981 HEAD^{tree}.

Filed, not folded in

Six non-blocking items both lanes raised, including whether the README should recommend ~> 0.2.0 rather than ~> 0.2: ~> 0.2 admits 0.3.0 and so spans the next break the way ~> 0.1 spanned this one. ~> 0.MINOR is a common Hex idiom, so that is a maintainer's call and is not made here.