feat(readme): publish the context-window cost as a badge, guarded by CI

[VickyXAI]

Stacked on #125. Base is perf/strip-json-schema-dialect, not main. The 12,900 figure assumes the $schema header is gone; without #125 every number here is 300 higher.

Why

A package manager has shown you install size for thirty years. Installing an MCP server spends your context window on every turn, whether or not you call the tools — and essentially no server tells you how much. There is no norm because nobody established one.

The precedent worth copying is bundlephobia: it did not win by adding a badge, it won by defining a metric and shipping the tool that measures it, after which bundle size became a competitive axis for everyone. Same move here.

What ships

A badge and a table. The hero gets 🧮 12.9K Context Tokens, and the profile section — where a reader is already choosing — gets what each profile costs:

Profile	Tools	Context
full (default)	20	12,900
trading	9	5,554
media	7	5,436
research	6	3,024
chat	3	1,924

--profile trading is 57% less context than the default for the same workflow. That was already true and already shipped; it was just never stated anywhere a user would see it.

The harness, committed. scripts/measure-tool-schema.mjs (npm run measure:schema) runs against this server or, with -- <cmd>, any stdio MCP server. The claim is checkable by anyone, against anyone — which is the part that makes it a standard rather than marketing.

A guard. test/schema-tokens.test.ts measures the live server in-process and fails when the README disagrees. It asserts the published form (12.9K) and the table's exact totals: pinning raw counts would fail CI on every wording tweak, while pinning the published figure fails precisely when the claim stops being true. Verified by deliberately corrupting the badge and a table row and confirming each fails.

Honesty, stated in the README rather than omitted

• Counts the model-visible projection ({name, description, input_schema} + host prefix); excludes annotations/_meta/outputSchema, a further ~3.7% on the wire.
• o200k_base is a proxy — Claude's tokenizer is not public and runs higher on JSON, so every figure is a slight under-count, never an over-count.
• Schemas are prompt-cached: 100% context cost every turn, roughly a tenth of that in dollars.
• 54% of our own cost is descriptions, not schemas. The badge does not flatter us; that work is not done.

Verification

• 429/429 tests pass (425 before, +4).
• Typecheck, build, and sync-brand-numbers --check all clean.
• The number reproduces to the token across four independent implementations: this script, an earlier Python harness, a third-party rewrite, and the in-process test.

Follow-up, not in this PR

Promoting mcp.contextTokens into brand-numbers.json needs a matching change in blockrun's src/app/brand/numbers.json/route.ts first — adding the key here alone would be dropped on the next --refresh and break markers. Kept out deliberately.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WUL3ExR4Nz7uebxKaKwjDi

[VickyXAI]

Superseded by #127 — same commit rebased onto main after #125 merged, plus the 0.45.0 bump. GitHub auto-closed this one when its base branch was deleted; it cannot be reopened with a deleted base.