docs(bench): CONTINUOUS_BATCHING.md cites a benchmark record that was never committed

## Summary

`docs/CONTINUOUS_BATCHING.md` names a benchmark record file that is not in the tree. The prose landed with the qwen3_moe batched-decode change, but the record it cites was never committed, so the reference is dangling.

## Background

The continuous-batching doc attributes a fallback measurement to a dated M1 Ultra profile. Readers following that citation reach a path that does not resolve. A dangling `docs/`-relative link is a small but real documentation defect: it implies data exists in the repo that does not.

## Proposed Solution

Either commit the missing benchmark record at the cited path, or reword the sentence to drop the file citation while keeping the measured claim attributed to the PR that introduced it.

## Implementation Notes

- `docs/CONTINUOUS_BATCHING.md:47` reads: "The 2026-09-04 M1 Ultra profile in `docs/benchmark_results/moe-batched-decode-m1ultra-2026-09-04.md` measured that fallback on `qwen3-30b-a3b-4bit` before the override landed."
- `docs/benchmark_results/` exists and holds sibling records (for example `autotuner-m1ultra-2026-07-30.md`, `cascade-attention-m1ultra-2026-08-03.md`), but not `moe-batched-decode-m1ultra-2026-09-04.md`.
- The sentence was introduced by commit `3a5c75d0` (PR #1623), which added the prose but not the record. This is the only dangling `docs/`-relative path found across `docs/*.md`, `README.md`, and `CONTRIBUTING.md`.
- Regenerating the record needs an M1 Ultra. If the original author still has the profile output, attaching it here turns the fix into a one-file PR (add the record). Otherwise the rewording path avoids the hardware dependency.

## Acceptance Criteria

- [ ] The citation in `docs/CONTINUOUS_BATCHING.md` resolves to a committed file, or the sentence no longer names a non-existent path
- [ ] The measured claim remains attributed (to the committed record or to PR #1623)

---

## Original Suggestion

### Title: docs(bench): CONTINUOUS_BATCHING.md cites a benchmark record that was never committed

`docs/CONTINUOUS_BATCHING.md` cites a benchmark record file that was never committed: the prose landed with the qwen3_moe batched-decode perf change (#1623), but the record it names is not in the tree.

## Evidence

- `docs/CONTINUOUS_BATCHING.md:47` — "in `docs/benchmark_results/moe-batched-decode-m1ultra-2026-09-04.md` measured ..."
- `docs/benchmark_results/` does not contain that file (introduced by commit `3a5c75d0`, PR #1623, which added the sentence but not the record)

This was the only dangling `docs/`-relative path across the checked prose in `docs/*.md`, `README.md`, and `CONTRIBUTING.md`.

## Suggested fix

Either commit the missing benchmark record or reword the sentence to drop the file citation while keeping the measured claim attributed to the PR.

**Caveat:** regenerating the record needs an M1 Ultra; if the original author still has the profile output, attaching it to this issue makes the fix a one-file PR for anyone.

## Acceptance criteria

- [ ] The citation in `docs/CONTINUOUS_BATCHING.md` resolves to a committed file, or the sentence no longer names a non-existent path

Start at docs/CONTINUOUS_BATCHING.md:47 and inspect docs/benchmark_results/ alongside the sibling records named in the issue. Check commit 3a5c75d0 and PR #1623 to confirm the measured claim's provenance; the work is done when the citation resolves to a committed record or no longer names the missing path, while preserving attribution.

Written by the indexing model from the issue text.