Exp Probe: ask once if needed → run → record → one line

A probe exists to answer one question cheaply. Everything about it is minimal except honesty.

• Apply goal-first: state what the result will change and when this check ends. Negative or inconclusive evidence can be useful; a surprising side finding does not silently expand the question.
• Define the single question first ("does config X fit in memory", "is this NaN reproducible", "what is the missing anchor value"). If you cannot name the question, this is not a probe.
• Shrink, don't distort: reduce sample counts and steps, never the semantics being measured (sequence length, batch semantics, dtype). State exactly what was shrunk when reporting.
• Match the question's path: integration and capacity probes use the real launcher/flags. A scientific tensor/layer proxy may replay a smaller real computation with its required inputs/state and checked reference output; it proves only that local claim, not end-to-end integration. Use exp-workflow for a proxy scheme that still needs co-design, then return here to execute the settled question.
• Probe code lives in the project's scratch/experiment area, isolated per probe: minimum code, real errors exit as-is; environment failures (device contention, disk full) are diagnosed as environment, not patched around. Do not add speculative defensive scaffolding.
• Preflight even when cheap: the same pass as a real run, one look — environment active, and a live device read with the platform's own tool (nvidia-smi, npu-smi, mthreads-smi, …) that picks a free device and passes it explicitly through the platform's visible-device variable (never the default). A probe that lands on someone else's device wastes their run and yours.
• Probes are meant to be cheap, so waiting inline is normal. If a probe's run is long (full-size model sanity check, slow dataset), use the same delegated pattern as exp-batch: background launch + later wake-up instead of holding the turn.
• One question to the user at most, and only if the question itself is ambiguous. No plan docs, no grilling rounds, no sub-agents.
• Record where the campaign can find it: scratch area per project rules, with config, observation, and exit status. If the probe's answer prunes campaign cells, the claim and its evidence belong in the campaign's pruning ledger — say which cells it affects.
• If the observation changes a major route or key understanding, use save-mem to link it into existing project records; otherwise keep the local probe record. One local observation does not settle every route or condition.
• Report in one line: question, observed answer, cost, and what it unblocks or rules out. If inconclusive, propose another probe only when it adds discriminating information worth its cost; otherwise state the limit and return to the main decision. Never silently upgrade into a batch.