docs: describe tilth by what it does, retire the cost claim

[jahala]

Summary

The README, the landing page and the package descriptions led with "−40% cost per correct answer". This PR removes every cost, accuracy and turn figure and describes tilth by what it does.

Why the number goes. It was measured in March 2026 on v0.5.0 with models two generations old, as point estimates with no interval. Regrouping our own 3,791 stored runs on matched tasks flips its sign by period. The grading audit of July 2026 found defects in the harness. No tilth-versus-baseline run exists on a current model. We have nothing measured to put in its place, so the README says that in a short note and makes no cost claim.

What replaces it. tilth parses code and computes answers that text search has to infer: where a symbol is defined, who calls it, what it calls, what depends on a file, and what changed at function level. Anyone can check that by running the tool. The limits are stated too: tilth parses syntax, does not resolve types, and matches callers by name.

Changes

• README.md: rewritten opening, a "What it doesn't do" section, a note on the retired benchmark. Every example is now real output from the FastAPI and Gin repositories, cut only at whole lines marked .... The old examples were hand-written and no longer matched the tool (faceted search sections, ASCII -> and -- calls --, the deps and grok layouts).
• index.html: the benchmark table becomes "questions grep can't answer", four rows of explanation plus real output. The hero shows a real callers query. Ligatures are off in code so -> and -- render as the tool prints them. The install grid no longer scrolls sideways on phones. The garden footer gains pollen and weeder, with accents from jahala/plotplot.
• Facts corrected: about 35,000 lines of Rust (was 20,000), 17 languages with Bash (was 16), and the claim that search time is constant regardless of repository size, which stopped being true for content search in fix(search): disclose the display cap, bound what content collects #220. The speed table is re-measured: median of 15 runs of the release build on an Apple M5 Pro, on Gin (130 files) and FastAPI (2,700 files).
• Cargo.toml, npm/package.json: descriptions drop "indexed" (tilth has no index) and gain a homepage. npm/README.md is added; package.json already listed it.
• benchmark/ removed (second commit). The harness is no longer run, and A/B tests moved to copeca. The last commit containing it is tagged benchmark-archive. The four pinned fixture repositories move to scripts/fixtures/setup_repos.py, a standalone script that re-pins with fetch and checkout and never deletes. CLAUDE.md, ARCHITECTURE.md, dependabot.yml, .gitignore and a Cargo profile comment drop their mentions.

Test plan

• cargo test: 680 passed
• cargo fmt --check
• scripts/regen-agents-md.sh is a no-op
• scripts/fixtures/setup_repos.py: fresh clone, idempotent rerun, unknown name (exit 2), re-pin of a clone on another commit
• Landing page checked in a browser at 1440 px and 390 px, light and dark: no horizontal overflow, no clipped lines in the output panes

Follow-ups, not in this PR

• The open garden PRs need reconciling: docs(garden): the bed's papers — garden.json, product layer, footer, metric summary #212 names cost per correct as the metric, and feat(core): split the parser into the tilth-core library crate #210's parity script should point at scripts/fixtures/setup_repos.py.
• skills/SKILL.md still shows ── calls ──; the tool prints -- calls --. feat(mcp): context diet — a pointer at initialize, the guide in skills/SKILL.md #211 rewrites that file.
• Outline signature lines double the comment marker for Go and Rust (// //), from src/read/outline/code.rs.

🤖 Generated with Claude Code