0.73.0

Added

• A kind: command step may declare expect: failure: a nonzero exit is then the passing verdict, and an exit of 0 submits failing evidence whose reason, nothing reproduced, is printed as a note of the run. A process killed by a signal is still no verdict, and so is an exit of 126 or 127 (a shell could not execute the command, or did not find it), with a reason naming the status; any other nonzero exit, 128 and above included, is a reproduction. expect: failure beside an evidence record: is refused at <step>.expect with unsupported_construct. A step without expect behaves, and digests, as before.
• aep drive run --input <name>=<value> (repeatable) gives a run named inputs, which a step names as {input.<name>} in its run words, its evidence record: path and an llm step's context. A run whose map names an input it was not given is refused before a run id is allocated, with undeclared_reference at the step naming the placeholder; a resume uses the inputs recorded in the run's launch.json. A value is filled in as text and never read again for placeholders, so two values cannot spell one between them; a value that names a placeholder, or that leaves an llm step's context entry blank or a glob, is refused before a run id is allocated.
• A command step may declare output: <file>: its standard output is written to {run_directory}/<file>. An llm step's context may name {run_directory}/<file>, and the session is handed that file. An output name that is blank, holds a path separator or a NUL byte, starts with ., is one of the run directory's own records in any ASCII case, or has the form of a step's log, <state>-<index>-<attempt>.log, is refused at <step>.output.
• The ESS specification (ess/domains/drive.yaml) models expect, output, run inputs, and reading a command's exit status against its expectation.
• aep plan artifact evidence <id> --kind task_spend --from <file> records what one source says a task spent, as a task_spend evidence record: model calls, the five token figures (input, output, cache read, cache write, thinking), a token total, cost in micro-USD, wall time, and for a driven run the llm steps declared and run. The file is a transcript (Claude Code stream-json or metaharness.event/1), an eval.run-manifest/1 (its tokens, cost_micro_usd and wall_time_ms) or a driven run's cursor.json. The token total has one rule for every source, the one a run manifest's tokens follows: input, output, cache read and cache write tokens, stated only when all four are known; thinking tokens are not added. Several terminal records of one session are charged once, as aep eval run charges them; across sessions a figure is stated only when every session states it and the sum fits, so a session that states no cost, or never reaches its terminal record, leaves the task's figure unknown. A figure the source does not state is left out of the record and printed unknown, never 0. The record keeps the source's sha256: digest and its path from the repository root with links resolved; a source outside the repository, a directory link out of it included, is refused with SpendSourceOutsideRepository, naming the path. A file no reader recognises as a transcript, manifest or cursor, or a total_cost_usd that is neither a number nor null, is refused with SpendSourceUnreadable, naming the file; --kind task_spend without --from is refused with SpendSourceDigestMissing. Each refusal writes nothing. --at may name when it was observed.