v0.4.0

0.4.0 (2026-08-07)

Features

• refuse to set a derived key instead of storing it and ignoring it (#64) (cea2d89)
• salvage the pure-Roc improvements from the reverted #69 (4df8fe9)
• score each activity at the FTP in force when it happened (a668cb5)
• stream activities.csv out of the export zip instead of extracting it (6320854)

Bug Fixes

• build the e2e harness with --opt=dev too (cbf0ef9)
• correct the cold-start date comparison and index the period-FTP lookup (0efe594)
• count unanalyzed activities in summary totals, and say when analyze stopped early (bee27ef)
• emit no speed sample across an unrecorded gap (f01c199)
• four scoring-accuracy corrections from the training-science audit (7b81c6d)
• handle both no-baseline cases in the compare screen, and use False for False (c177e6e)
• interpolate cumulative streams so coarse recordings don't spike NGP (df91559)
• point the justfile at the new compiler so just test works locally (7402ca3)
• reject pause-spanning windows when deriving bests (fc08e4b)
• say when CTL is still warming up, and pin the load model over many days (f39bf58)
• show the warming-up warning to humans, not only to JSON callers (0ad2304)
• sort stream samples before resampling, and stop dropping light sessions (cda9c8f)
• stop plan showing duplicate rows for a re-planned day you then missed (a87b752)
• stop counting coasting as easy riding (97f369f)
• sum real seconds in the pace intensity split, and call roc by name (3271606)
• use the real EWMA factor so 42 and 7 days mean 42 and 7 days (8d18965)
• warn about an unconverged CTL until 90 days, not 42 (35681d5)