handbook: a step that produces nothing still reports success

[inhuman]

Three classes from a week of live work, none of them in the handbook.

F9 — the silent no-op. A skipped when, a call error swallowed by
on_error: continue, a variable resolving to empty: all of them reach the trace
as successful steps with non-zero duration. One pass was placed four times
before it applied anything, and every trace looked identical — all steps fine,
turn completed. The four causes were a wrong argument name, a gate testing for
the presence of success instead of the absence of failure, a step sitting
outside the branch that owns its variable, and a neighbouring loop wiping that
variable on every clean run. All four were found by comparing a FACT with an
expectation, never by reading the trace.

F10 — a branch that never fired is dead, not rare. Measured on a live
catalogue: the only dead branches sat in one description, where half the
branches were never chosen and two fallbacks did the work. The threshold
matters — a rule of "never fired in twenty runs" missed it, because dead
branches live where traffic is thin.

Context section — an extractor whose output the model reads must be
bounded, because it grows with the data. A cluster-wide list was printed in
full and the next step died with a context-length error. Volume has to be made
artificially: stored fixtures are trimmed, or they weigh too much.

Verification, V7 — code inside steps is checked by none of the first three
layers: the linter does not execute it, and the step probe visits only steps
with a schema. Two cheap layers close the gap, along with the two traps that
made each of them pass on any code.

The package's own test caught a mistake in this text — it named a trace field as
if it belonged to the format. Reworded.

🤖 Generated with Claude Code