fix(graph): take the benchmark's own symbols out of the tour's schema

[samchon]

The leak

The reinterpretations field illustrated itself with

`["track", "trigger", "ReactiveEffect", "setupRenderEffect", "queueJob", "patch"]`

Six Vue internals — and Vue is one of the eight repositories this benchmark measures. Every model copied all six, verbatim, into its Vue answer:

model	names it sent	copied from the example
Sonnet 5	9	6 of 6
Opus 4.8	19	6 of 6
GPT-5.6 terra	8	6 of 6
GPT-5.6 sol	13	6 of 6

A schema example that names a measured repository's symbols is the tool handing one cell its answer, and it buys a private codebase nothing. The example is now four names from no codebase at all — JobQueue.push, Scheduler.tick, Worker.run, drainQueue — and the field's prose names no repository either.

What it cost

Nothing, as it turns out — which is the only reason the published grid was not wrong. Re-measured leak-free, Vue holds at 92% and one call for Sonnet, exactly what it scored with the example spelling out its answer. The models name track and trigger from the question, not from my example. What moved is everything else:

axis	before	after
Sonnet / common	92%	92%
Sonnet / dedicated	87%	92%
Opus 4.8 / common	93%	93%
Opus 4.8 / dedicated	91%	91%

Sonnet's project-question axis gains five points, with seven of its eight cells improving: excalidraw 92% (5 calls) → 98% (1), zod 87% (2) → 92% (1), typeorm 85% (2) → 90% (1), NestJS 82% (3) → 88% (2), shopping-backend 77% (3) → 85% (2), VS Code 83% (2) → 87% (1). 0 file reads, 0 shell commands.

The Codex axes could not be re-measured on this build (no quota) and keep their previous cells.

Three more things the schema was asking for and should not have been

• The results echoed the question back. ITtscGraphTour.query and ITtscGraphEntrypoints.query returned the caller's own string to it — the string it wrote two lines above, in the same call.
• The MCP instruction listed reinterpretations among the chain-of-thought steps, beside question, draft, review and request. It is a field of one request branch, not a step of the reasoning. It now sits with the tour line in the request menu, where a branch's own fields belong.
• The field's description carried a chronicle of how it came to exist — how many cells a duplicate field cost, which model dropped which backtick. That belongs in a commit message, not in a schema re-sent on every turn. The tour branch's schema prose: 2,661 characters → 1,461.

Test plan

• tsc --noEmit and pnpm -F @ttsc/graph build clean; tests/test-graph unchanged and passing in CI.
• Benchmark: all 32 Claude cells re-measured on this build, plus one re-measurement each for the three cells that came back anomalous (Sonnet shopping-backend at 13 calls, Sonnet rxjs, Opus NestJS and shopping-backend — all model spirals, all normal on the second run).