# Reproducible benchmark evidence

## Command and environment

```sh
make benchmark \
  EMACS=/Applications/Emacs.app/Contents/MacOS/Emacs \
  LOAD_EXTRA='-L /tmp/tp-elpa.Ur243x/dash-2.20.0'
```

- Emacs: 30.2
- Fixed seeds: `1`, `7`, `42`, `747555`
- Reproducible generated seed: `8675309`
- Every scenario performs a correctness assertion before accepting timing.
- The full command completed successfully on 2026-07-28.

The runner is `tp-benchmark.el`; `make benchmark` is the stable entry point. Output rows include scenario, status, fixture, seed, requested/actual size, operations, scanned/changed/refreshed counts, elapsed wall time, and GC count.

## Representative result: seed 42

| Scenario | Requested / actual | Operations | Scanned | Changed | Refreshed | Elapsed (s) | GC |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| large text | 100,000 | 2 | 100,000 | 100,000 | 0 | 0.000304 | 0 |
| large text | 1,000,000 | 2 | 1,000,000 | 1,000,000 | 0 | 0.000323 | 0 |
| fragmented intervals | 1,000 | 1 | 1,000 | 500 | 0 | 0.044254 | 0 |
| fragmented intervals | 10,000 | 1 | 10,000 | 5,000 | 0 | 0.404836 | 1 |
| fragmented intervals | 50,000 | 1 | 50,000 | 25,000 | 0 | 1.965066 | 3 |
| stack depth | 1 | 2 | 2,000 | 2,000 | 0 | 0.000909 | 0 |
| stack depth | 5 | 6 | 2,000 | 2,000 | 0 | 0.003204 | 0 |
| stack depth | 20 | 21 | 2,000 | 2,000 | 0 | 0.012905 | 0 |
| stack depth | 50 | 51 | 2,000 | 2,000 | 0 | 0.037830 | 0 |
| reactive fan-out | 1 / 1 | 1 | 1 | 1 | 1 | 0.000763 | 0 |
| reactive fan-out | 10 / 10 | 1 | 10 | 10 | 10 | 0.005469 | 0 |
| reactive fan-out | 100 / 100 | 1 | 100 | 100 | 100 | 0.051376 | 0 |
| reactive fan-out | 500 / 200 | 1 | 200 | 200 | 200 | 0.102989 | 0 |
| theme refresh | 1 / 1 | 1 | 1,000 | 0 | 1 | 0.001317 | 0 |

The 500-buffer fan-out fixture is intentionally capped at 200 live buffers on local hardware; output reports both requested and actual counts. Theme refresh uses the v1 conservative managed-range strategy.

## Interpretation

These figures are advisory, not release thresholds. Fragmented interval count is the clearest scaling risk and should remain the first optimization target. Large-text lookup over one property run and deep-stack top lookup are cheap in this fixture, but those results must not be generalized to fragmented text or repeated stack mutation.