[DanielLuning]

What changed

• New example examples/numeric_loader.exs: a numeric CSV (rows of -?[0-9]+\.[0-9]+ fields) and a flat JSON array of numbers, loaded straight into packed [f64]. Two passes: a vectorised count of newlines or commas, then one state-machine parse into an alloc of that size. The grammar is written down: at most 18 digits and a mantissa of at most 2^53, so one division gives the bits of String.to_float/1. Everything else raises ArgumentError.
• Elixir variants, all timed in each run: NimbleCSV + String.to_float, :binary.split + binary_to_float, :binary.match + binary_to_float (header only), and a binary-matching parser with the kernel's algorithm (the baseline). JSON: Jason, OTP's :json, and the same binary-matching parser.
• examples/scoreboard.json: the new row, quiet, stamped at 0d6e7c7 (--snapshot --fresh from a clean tree). .gitleaksignore: the sort_u64_keys pin moves from line 34 to 35, as the new row sits above it.
• examples/_tables.exs, both READMEs (regenerated), an example bullet, and a note in the wrapper table's parser row: a packed numeric load is an exception like BPE ids.
• docs/performance.md 2.18: what the Elixir side tried, the cost of the count pass, and the wrap64 count that vectorises.

Results

	best Elixir	defk	speedup	verdict
CSV, 26 MB (400k rows x 8)	138.8-143.7 ms	50.5-53.0 ms	2.68-2.78x	good
JSON, 2.2 MB (153,600 floats)	9.0-12.4 ms	2.08-2.34 ms	3.90-5.87x	good

• Eight runs. The stored run passed the quiet test for both reports. The CSV numbers of the other runs are provisional (up to 1.9 cores of other work after the measurement).
• The JSON Elixir side is bimodal: its best of 15 is 9-12 ms, its median 18-20 ms, depending on whether a run escapes GC.
• The count pass is 6% (CSV) and 11% (JSON) of the kernel's time, under the ticket's 20% line, so no growable output construct is needed.
• The binary-matching Elixir parser that appends to one binary is 5.7-5.9x NimbleCSV + String.to_float; with an iolist it was 1.7-1.9x slower.
• One kernel shape was tried and dropped: a loop over the fields with an Enum.reduce_while per field. It was 15% slower.

How it was checked

• mix ci green in the worktree on the Mac (2305 passed). test/scoreboard_doc_test.exs checks the generated tables.

• Agreement: every Elixir side, the kernels and their Elixir mode give the same bytes on 26 CSV and 28 JSON inputs. 411 edge fields are bit-exact with String.to_float/1. 23 bad CSV and 26 bad JSON inputs raise on every side that has the grammar.

• mix defk.examples --stale lists the new row as current.

• Touches a promised surface (kernel subset, runtime contract, public API, error codes)? Then a CHANGELOG.md entry, and a deprecation for a rename or removal (docs/guides/stability.md).

• Touches test/stability/? A frozen file is never edited; a release adds a directory (docs/agents/stability.md).