docs: honest results framing, CODEOWNERS, and a stale eval fixture

[theDakshJaitly]

Three separable commits, all following from the findings in #115.

1. README results framing

The Results table led with 916.38x and 10.74x under a bare "Results" heading. Both numbers are accurate for what they compute, but neither measures what a reader assumes.

The denominator in evaluate/lib/grep-baseline.mjs is ceil(chars/4) over the entire contents of grep's top-3 files, and no agent runs in that harness at all. It is a retrieval-payload ratio against a favourable baseline. evaluate/README.md already states plainly that the harness cannot support an end-to-end graph-versus-no-graph claim; the top-level README did not carry that across.

Independent measurement on a 1,600-file Rails codebase found mex spending 1.3-1.7x grep's total session tokens across three real tasks. A smaller payload per call does not recover the cost of extra round-trips, each of which re-sends full context.

Changes:

• Lead with the correctness result — 23/23 call sites found where a \.method_name regex found 15/23, the missing eight being bare self-calls with no explicit receiver. That is the claim the data supports, and it is the stronger argument.
• State the denominator inline in the table so the ratio cannot be misread.
• Replace the soft closing caveat with the measured 1.3-1.7x finding and a pointer to the three-arm experiment evaluate/README.md specifies.

2. CODEOWNERS

Contributors had no way to tell who reviews what, and were asking directly in issue threads. Starts minimal with a single default owner. The header comment records the three rules that make CODEOWNERS silently do nothing: last match wins, owners need write access, and it is advisory unless branch protection requires it.

3. Stale eval fixture

evaluate/fixtures/nl-tasks.json expected BudgetedEmitter for the budget task. That symbol does not exist anywhere in src/ and appears only in the fixture. The task was unpassable, so every natural-language retrieval measurement — including ones taken before this branch — was scored against a target that could not be hit.

The budget is enforced by BudgetLedger (src/graph/agent-protocol.ts, constructed in cli-agent.ts#beginResponse). Correcting it raises measured baseline recall on the NL fixture set from 3/7 to 4/7 with no code change.

Scope

Documentation, config, and one test fixture. No source changes.

Verification

npm run build clean, npx vitest run 369/369 passing.

[Copilot]

Pull request overview

Updates project docs and evaluation fixtures to more accurately represent what the current benchmarks measure, clarifies review ownership, and fixes an unpassable NL eval expectation.

Changes:

• Reframe the top-level README “Results” section to lead with the demonstrated correctness win, clarify the retrieval-size denominators inline, and explicitly state the lack of end-to-end token-savings evidence.
• Add a minimal .github/CODEOWNERS with a default owner and a brief “how CODEOWNERS can be ignored” header note.
• Fix evaluate/fixtures/nl-tasks.json to expect the real budget-enforcement symbol (BudgetLedger) instead of a nonexistent one.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
README.md	Reframes benchmark claims and clarifies what the reported ratios and harness do/do not measure.
evaluate/fixtures/nl-tasks.json	Updates the budget task expected/rubric symbol to match the implementation (BudgetLedger).
.github/CODEOWNERS	Introduces a default code owner and documents key CODEOWNERS matching/eligibility caveats.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.