[timoncool]

A planned score the semantic stage cannot follow is caught before the long stage and planned again with another seed.

• src/score-check.h: a plan is refused when its header has no well-formed M:, L: or K:, when it has no voice with bars, or when colons stand inside the note runs (outside chord symbols, inline fields and repeat bars). Adapters pushed past their limits or a seed that goes astray write such plans, and the semantic stage then spends minutes on one. The log shows the start of the score in printable ASCII and what is wrong; yue-plan only warns. tests/test-score-check.cpp covers the cases.
• plan_retries (request, default 2, bound 8): a refused plan of song i is planned again with lm_seed + i + k * lm_batch_size for attempt k, a seed no other song of the batch draws, so the same request replays to the same score and the replay request of the track carries the score that was sung. Once the retries are spent the request stops with a FATAL naming the problem. Outside the bound is a 400 from the server and a FATAL from the CLI.

Tested on an RTX 4090 (Q8_0 backbone, tests/request0.json style and lyrics, 30 s, seed 1234): test-score-check passes; a normal request plans 739 tokens and renders as before; with the score budget cut to 20 tokens every plan lacks K:, the log shows both retries on seeds 1235 and 1236 and then the FATAL; plan_retries: 9 is refused.