docs(benchmarks): replace unverified claims with measured numbers

Two claims had no measurement behind them. Both now do.

Incremental update latency
- "under 2 seconds on a ~2,900-file repo" was never benchmarked: there is no
  incremental benchmark in eval/benchmarks/ and no eval repo is near that size.
  Measured on django (2,927 .py files; graph indexes 2,998 files, 46,683 nodes)
  on an M4 Pro: cold build 40.3s, no-op update 1.4s, two-file edit 2.4-2.9s on
  the --skip-flows path the hooks use, 9.8s with full post-processing. The
  original claim holds only for the no-op case, so README/FAQ/TROUBLESHOOTING/
  LLM-OPTIMIZED-REFERENCE and diagram 4 now quote the measured figure, and
  docs/REPRODUCING.md gains an "Incremental update latency" section with the
  corpus, machine and commands.

Token benchmark re-capture
- The fastapi row carried a caveat instead of a measurement. Re-running from
  clean clones at the pinned SHAs showed the whole table was stale, not just
  fastapi: the local eval checkouts had drifted (fastapi alone had 2,219 dirty
  entries), so nobody could reproduce the published numbers from them.
- Re-captured all six repos: range is now 36x - 376x with a ~65x median, down
  from 38x - 528x / ~82x. avg graph_tokens rose in every repo because per-node
  embedding text grew richer; naive_corpus_tokens is unchanged for
  code-review-graph and gin, which confirms the movement is on the graph
  response side. fastapi is now measured at its current pin 22381558.
- Build stats re-captured for nodes/edges/embeddings; flow, community and FTS
  counts are marked "-" rather than carried over from the older run.

Diagrams 1, 4, 5 and 6 and the alt text in all five READMEs updated to match.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JNv8JqBb46stATZYUinQtn