[noderaven]

Release 1.6.6: short task-checks command.

On e1a, 7 of 40 task-checks calls never ran: the Bash tool on Windows fails a command of about 7,000 characters or more, and those commands were 6,907 characters or longer. The command carried every path in the task's Files list (--declared) and the lane's check commands twice (--lane-cmd and --wide-cmd), so it grew with the task.

Now setup's owner map (owners.json) records each task's declared Files list, and the engine passes a fingerprint of it instead: --declared-from-owners N:HHHHHHHH, the path count and a 32-bit FNV-1a hash of the paths. task-checks reads the list from the owner map and checks it against the fingerprint; on a mismatch it writes an error record and exits 3, so a wrong list never decides the check policy. When the wide list equals the lane list the engine passes --wide-as-lane instead of repeating the commands. The command no longer grows with the Files list (1 file and 200 files differ by at most 2 characters), and an e1a-sized unit's command is under 4,000 characters.

Also in this release:

• start-task --scope takes --declared-from-owners with --owners the same way; when it cannot build the list from the owner map it says so in the scope section instead of guessing.
• The agent rules name the limit: the Bash tool fails on a command of about 7,000 characters, a long heredoc included; to add to an existing file such as a report, Read it and then Edit it, or Write the whole file. The fix prompts say "add ... with your file-editing tools" instead of "append".
• The dry run warns when a unit's task-checks command or the integration checks command is over 6,000 characters, and suggests shortening the project's check commands, for example with a wrapper script.
• --declared and --wide-cmd still work for a call made by hand.

Testing: built by a parallel-lanes run on 1.6.5 (r166): 7 tasks in a prelude, 3 lanes and a join in two waves, each task implemented and reviewed, then integration, a three-lens final review (7 minor findings, 5 fixed; 2 declined, both about the engine's own join-wave merge messages, which have no conventional prefix: left for a later release) and a verify run. Fingerprint fixtures are shared by the Python and JavaScript tests, including quotes, spaces, non-ASCII paths and a lone surrogate. An end-to-end test on real git checks that the engine-built command gives the same check policy as the old long form. Full suite on Windows (Python 3.15): 1715 tests, 1699 pass, 16 skipped, 0 fail. Not run: a validation run on a real project.