close the mercury fast-path eligibility gaps. each is a MFP_BAIL in mercury_fastpath.cpp. stream input and resolution reduction are done.
layer limits: plan-stage packet filtering.
palette / ICC / channel definition: apply grok's existing post-processing to mercury's output rows on the host side, do not port into mercury.
region decode: crop rows in the stripe flow, decode full width, crop columns at output. alternatively classic keeps regions permanently.
PLT/TLM packet-length shortcuts in the planner: PLT pays off for packets the plan drops (reduce, layers, later regions), TLM replaces the sequential SOT scan. both are opt-in markers, test files need write_plt/write_tlm.
precinct smaller than code-block: rejected at plan stage, effective block-size clamping not wired.
part 2 transforms (DFS, ATK): bail to classic, and classic only decodes them whole tile at full resolution.
close the mercury fast-path eligibility gaps. each is a MFP_BAIL in mercury_fastpath.cpp. stream input and resolution reduction are done.
layer limits: plan-stage packet filtering.
palette / ICC / channel definition: apply grok's existing post-processing to mercury's output rows on the host side, do not port into mercury.
region decode: crop rows in the stripe flow, decode full width, crop columns at output. alternatively classic keeps regions permanently.
PLT/TLM packet-length shortcuts in the planner: PLT pays off for packets the plan drops (reduce, layers, later regions), TLM replaces the sequential SOT scan. both are opt-in markers, test files need write_plt/write_tlm.
precinct smaller than code-block: rejected at plan stage, effective block-size clamping not wired.
part 2 transforms (DFS, ATK): bail to classic, and classic only decodes them whole tile at full resolution.