test(mcp): replace exact context-bytes pins with tolerance bands

[bokelley]

The exact-equality pins on standalone.context_bytes and the dictionary-definition counts in tests/mcp-schema-analysis.test.cjs force every schema-touching PR to re-pin the constants — guaranteeing merge conflicts between any two in-flight schema PRs. PR #6279 was knocked from MERGEABLE back to CONFLICTING four times in two days by exactly this line.

This replaces the pins with the invariants that matter:

• standalone context within [200 KiB, 400 KiB] (large enough for the experiment comparisons to mean something, below the projection budget)
• shared dictionary stays substantial (>100 definitions)
• prompt cleanup strictly reduces definitions

The ratio comparisons (cleanup < 0.82×, shared dictionary < 0.37×) and the payload-consistency equality are unchanged — they carry the experiment's real signal.

Closes the snapshot-pin half of #6571; the budget-headroom half stays open pending the shared-dictionary graduation.

🤖 Generated with Claude Code

[aao-secretariat]

Ladon verdict: Approve

Approve — clean test-hygiene change.

This PR touches only tests/mcp-schema-analysis.test.cjs (+20/-5), replacing three brittle exact-equality snapshot pins (context_bytes == 287_357, dictionary_definitions == 148 and == 127) with a tolerance band plus strict-reduction inequalities. Ratio comparisons and the payload-consistency self-check are untouched, so context_bytes is still validated against its computed source rather than a frozen literal. No published protocol surface (schemas, docs/reference, dist artifacts, task defs) is touched, so no changeset is required. Well-justified fix for real merge-conflict churn (#6279, #6571).

Decision-table walk: no critical/high findings (row 1 no); gated_paths false (row 2 no); high_risk false, no deletions/modifications-with-medium (rows 3–5 no); no prior decision (row 6 no); no no-auto-approve team match (row 7 no); zero medium findings (row 8 no). Falls through to row 9 → approve.

Nit (non-blocking): line 18 lower-bound literal 200_000 is mislabeled as "200 KiB" in the error message.