docs: replace the invented examples with measured ones

[yanxue06]

Three of the README's four code samples are fabricated. The demo at the top shows "Alice Chen" and "Bob Martinez" committing to a repository that does not exist; the JSON sample reports 48,213 candidates; the tuning section prints a search over the same imaginary history.

A reader evaluating the tool cannot tell any of it from real output — and the real output is more convincing than the invented output was.

Measured, not invented

Everything now comes from runs against a fresh clone of ripgrep, 2,287 commits.

The demo is git log --grep next to the same question asked in English:

$ git log --grep="large files"
139f186 crates/ignore: switch to depth first traversal
714ae82 Add `--max-filesize` option to cli
d06f84c Get rid of special mmap decision on Windows.

$ git-semantic search "searching very large files without loading them into memory"
1. ca058d7 - Add support for memory maps. (0.75 similarity)

Grep returns three commits, none of them the one that answers the question.

Index and query cost are measured too — 16 s to build quick, 83 s full, 4.1 MB and 7.8 MB on disk, 100 ms per query end to end on an M5 Pro. The cargo bench table stays: it measures retrieval in isolation, which is why its numbers are three orders of magnitude smaller, and the two figures now explain the difference instead of quietly disagreeing.

Documents this release

• Quick Start is one command, now that search bootstraps itself (feat(cli): search bootstraps itself on first use #84). init and index move down to "if you want to do it ahead of time".
• git semantic search … works, and has since the day the binary was named git-semantic. The README never said so.
• completions (feat(cli): shell completions #83), --before, and --force were undocumented.
• update is deprecated and no longer listed.
• Subdirectories, worktrees, and submodules get a mention, since until fix(index): find the git dir from any directory in the repository #82 only the repository root worked.

Two corrections

Rust 1.70+ is wrong. The crate is edition 2024 and uses let-chains throughout the index and search paths, so 1.88 is the floor — verified with cargo +1.88 check --all-targets, and now declared as rust-version = "1.88" so cargo reports it instead of failing on a parse error. crates.io shows the MSRV on the crate page as a side effect.

Sub-millisecond retrieval was measuring the graph traversal while the CLI prints a figure that also includes embedding the query. Both were true; together they read as a contradiction.

The LICENSE badge also now links absolutely — crates.io renders this same file and cannot follow a relative path to LICENSE.

---

Structure and prose are otherwise untouched: the RRF-versus-weighted-blend rationale, the MMR explanation, the tokenizer notes, and the benchmark table are all kept as they were. Diff is +105/−44.

Last of five PRs, on top of #85.