[guijacquemet]

Fixes the first theme of the round-trip findings. Every case that the matrix skipped for these findings now runs as a normal test; the findings are removed from tests/roundtrip_known_failures.py.

Before -> after

• Numbers (F1): an integer or float given text, a list, an object, NaN, infinity or a huge int failed with the Python exception name as code; now invalid_parameter ('N' must be an integer, got <what>). 3.0 is still accepted for an integer; the text "3" is refused for a number (hosts must send a JSON number).
• Paths (F12): a non-text file, folder or image value is invalid_parameter.
• Shared memory (F11): a stale or malformed descriptor is unreadable_image with a message saying what the host must do.
• Wrong returns (F5): a wrong kind of returned value is bad_return naming the output and what was expected; a message output that is not text, a file output that is not a path, a values output that is not a mapping, and several outputs given a bare array are refused too (before: str() of whatever came back, or a numpy truth-value error).
• Affine (F6): NaN or infinity is bad_return (was delivered as null).
• Image outputs (F15, F20): empty arrays are bad_return; complex, object, text and bytes arrays are unsupported_dtype; uint64 beyond float32 is refused with unsupported_dtype on every machine (the float32 exactness check no longer uses an out-of-range cast, which stored 264-1 as 264 on ARM).
• Hint misuse (F16): Min/Max on a non-number, a non-finite bound, Min above Max, PixelSizeOf on a non-float, Axes on a non-image, Name/Replace/ApplyTo on an input, and two outputs with the same Name are now a DeclarationError at registration (they were silently accepted or renamed). An app that carried one of these mistakes will stop registering: I described the tools of VLab4Mic, Playground, GuessTheCondition and NucleiSky (bridge-test) with this branch and all register; CellTracksColab was not checked.

Test edits (exceptions to "no existing test edited"): tests/test_sabotage.py: four break patches changed text because the lines they patch moved (same breaks). test_roundtrip_known_failures.py::test_the_skip_is_exactly_the_list used an F1 id as a sentinel; replaced by a check independent of any one id. No existing test encoded the buggy behaviour.

Docs: PROTOCOL.md (value rules, returns, image outputs, uint64), MANIFEST.md (refused hint misuse), REGRESSION_LEDGER.md.

Host follow-ups (not in this PR): mirror the messages above where a host validates before sending, send numbers as numbers and paths as text, send the full shared-memory descriptor and keep the block alive until the task ends. Tracked in the host parity work.

Verified by the author: black, ruff, mypy, vulture; the full suite has only the two environmental failures of this sandbox. Not verified: Windows, macOS, ARM, the coverage gate (CI runs it).

🤖 Generated with Claude Code

https://claude.ai/code/session_01Y2SN8SuEakTdiWy3F2WUqN

---

Generated by Claude Code