[DanielLuning]

Merge #74, then #75, first. This branch builds on defk-9p6.30/tape-csv (the same tables, README rows and scoreboard).

What changed

• New example examples/threshold_cross.exs: the first packet whose running total passes a byte budget. The loop is sequential, stops early and is integer-exact. It is written four ways: Elixir (E1, E2 with four packets per clause), a kernel (Enum.reduce_while), defn with while and a stop flag, and defn with cumulative_sum + argmax. EXLA runs on resident tensors, one Nx.Defn.compile/3 per shape.
• The example also measures per-call times from 16 to 16M packets, 1 and 48 concurrent callers, the tick gap on one scheduler, and 300 calls of changing length (EXLA's jit cold and warm, and the power-of-two bucket workaround).
• docs/scope.md: the sentence "not measured against EXLA yet" is replaced by the numbers. docs/research/nx-boundary.md: the open question is now a list of measured numbers. docs/performance.md 3.10. _tables.exs, the READMEs (regenerated), an example bullet.
• examples/scoreboard.json: a quiet row at 3602f17 (--fresh from a clean tree).

Results

Per call, µs, with the budget passed at 25%. Two runs; the first shared the machine with a mix ci, so these are provisional.

n	E2	defk	EXLA while	EXLA scan
16	0.14-0.16	0.05	40-48	37-126
4096	1.03-1.09	0.36-0.42	40-41	50
64k	15.7-17.0	4.4-4.6	62-63	154-168
16M	4336-4467	1473-1478	4272-4359	24780-25650

• Verdict (64k packets, against Elixir): 2.5-3.4x, good. EXLA while is 11.7-13.2x slower than the kernel.
• Kernel against EXLA: the kernel is faster at every size, from about 790x at 16 packets to 2.9x at 16M.
• Elixir against EXLA: optimised Elixir beats EXLA while up to 1M packets and is level at 16M. The cumulative_sum form is slower than Elixir at every size.
• Concurrency (4096 packets, calls per second): EXLA goes from 24.5k (1 process) to 46.8k (48 processes); the kernel from 3.2M to 22.3M.
• Variable length: EXLA compiles each new length, about 6.4 ms each (1.9 s cold, 20 ms warm). Buckets bring it to 0.25 s cold and 36 ms warm. Elixir takes 0.8-1.0 ms, the kernel 0.4 ms.
• Tick gap: about 2 ms on every side. EXLA's dirty NIF does not hold a normal scheduler.

How it was checked

• mix ci green in the worktree (2305 passed).

• Agreement: all sides give the same index on 80 inputs, including the kernel's Elixir mode. An empty input works in the kernel; Nx.from_binary(<<>>) raises.

• 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).