v0.25.0

[0.25.0] - 2026-09-25

Added

• Test Flow authoring — the Test Flow panel (View > Test Flow) no longer only runs a hand-written flow; it writes one, closing issue #155. Pick a component and one of its real serialize() keys from a combo that lists every key with its JSON type and current value, seed the values from the live circuit, add/update/remove conditions and measurements, Save Flow... it, and Reload to pick up a hand edit without re-picking the file. New from Circuit scaffolds a draft from the live circuit — a measurement on the first component with an output port, port 0, power_dBm, the form pre-filled from the first sweepable key and its current value — so a working file is one Add-condition click away, and every id and path in the form is read from the circuit rather than invented. A non-numeric key is listed but disabled with its JSON type as the reason, and a condition whose path does not resolve is marked red and answered with the sweepable keys that engine's serialize() does expose (up to kMaxPathHints), because a mistyped serialize() key, never an inspector label, is the likeliest authoring mistake.

Changed

• Flow authoring rules live in the harness, not the panel — test_flow/ gains describeConditionPaths() (every scalar leaf of a component's serialize() snapshot with its JSON type and current value, containers descended and never listed; only the entries it marks numeric are paths resolveConditionSlot() accepts, so a picker built on it cannot offer a path applyConditionValue() would refuse), buildFlowDocument()/writeFlowFile() (the document LoadFlowFile() reads back, ids and paths verbatim, written atomically via temp + rename so a failed write cannot truncate a hand-authored flow), and parseConditionValues()/formatConditionValues() (the values field's grammar and its shortest round-trip inverse, refusing rather than skipping a token that is not a complete finite number). The panel keeps adding no second copy of the document shape, value grammar, or applicability rules, and the harness keeps its no-ImGui/no-simulator::app rule.
• A draft is the single thing edited, validated and run — edits land in an in-tool draft and spec() is the one accessor preview(), run() and saveFlow() read, so the panel cannot validate one flow and run another; addCondition()/updateCondition()/addMeasurement() check each candidate through ValidateFlow() and report its own wording, plus the loader's duplicate-target/duplicate-reading rules that ValidateFlow() does not check, so a refusal at the form is the verdict Run would give later and a draft the panel runs is also a file that loads back. A load clears the draft (the file is the truth again); a circuit reload keeps it and preview() revalidates it against the replacement circuit.
• app/AGENTS.md and test_flow/AGENTS.md document the authoring contract: the spec()/draft boundary, circuit-sourced references, the form's use of ValidateFlow(), and the values-field buffer growth.

Fixed

• A refused authoring edit no longer creates a draft — the duplicate-target and duplicate-reading checks and the remove/update index checks now run against the effective spec before any draft is materialised, so draftActive() keeps meaning "unsaved edits exist" and a rejected click cannot flip the panel into that state.
• The values field no longer truncates a long sweep — its ImGui buffer grows from the model text (kValuesBufferHeadroom past its end) instead of using a fixed 256 bytes, which had let a long sweep's seeded text be truncated on the next keystroke and the truncation stored back, silently dropping values from a list the harness would then run as if it were complete.
• Discarding unsaved edits is stated, not silent — a load (and reloadFlow(), which goes through it) reports in the status that the in-tool edits were discarded, and New from Circuit reports when it replaced a draft.

Testing

• Add four [issue155] cases to tests/test_issue87_flow.cpp: the offered paths are exactly the sweepable leaves of serialize() (discovery and resolver agree, including that resolving touches nothing), a built document is the flow LoadFlowFile() reads back, a mistyped value list is refused rather than partially accepted, and a failed write leaves the previous file byte-identical.
• Extend tests/test_test_flow_widget.cpp with the authoring contract over real headless frames and clicks: a flow scaffolded from the circuit saves, loads back and runs (with registry-id pinning), an authored condition is refused exactly where the harness would refuse it, the form seeds the current value and the draft is what runs, reload re-reads the selected file and drops the draft, a circuit reload keeps the draft and revalidates it, a scaffold needs something to measure, a refused edit never creates a draft, a 200-value list survives the form round trip, and the authoring buttons drive their own actions.