[Lukas-Dresel]

decompile-project --stream --jobs N is the only way to get live progress (.streaming, index.jsonl), but it also writes the .c as functions finish, in worker completion order, so two runs of the same command differ. A pipeline that stores the export content-addressed has to choose between progress and reproducible output. --progress gives the progress over an ordinary export:

kuna decompile-project /usr/bin/ssh -o plain --jobs 4
kuna decompile-project /usr/bin/ssh -o p4 --progress --jobs 4
kuna decompile-project /usr/bin/ssh -o p2 --progress --jobs 2
diff -rq plain p4; diff -rq p4 p2
# Only in p4: index.jsonl

While it runs, .streaming reports as it does under --stream, from the load on:

{"schema":1,"phase":"decompiling",...,"jobs":4,"seeds":0,"functions_total":1090,"functions_done":173,"functions_failed":0,"seconds_since_last_result":0,"c_bytes":0,"asm":"pending","error":null}

	progress feed	artifacts written	reproducible under --jobs N
neither flag	no	at the end	yes
--progress	yes	at the end	yes
--stream	yes	as the run goes, decompile order	no

The change:

• --progress runs the non-stream export unchanged (same plan, structure naming, merge and artifacts) and reports on it. .c/.h/.asm/README.md are byte-identical to the export without the flag.
• index.jsonl is a feed while the run lasts: one line per function's first result, in completion order, with c_offset/c_len null. At the end it is replaced by the address-ordered index of the final .c, then .streaming is removed.
• The previous export's .c/.h/.asm are removed once the load succeeds, so a stale body never sits next to a new feed.
• --stream is unchanged and wins when both flags are given. --progress accepts --assert under the non-stream rules.
• Hooks: jobs::PoolWatch on the pool's first pass, a landed callback on the serial and callee-first loops, ExportWatch around the export.

Tests: a_progress_export_is_the_non_stream_export_plus_its_index (byte-identity at --jobs 1/2/3), a_progress_export_reports_progress_before_it_writes_the_export (reads the feed mid-run during an injected worker stall), stream_wins_over_progress, and a README render test. Datatests and stage tests unchanged.

One limit, inherited from the non-stream export: a function that hits --max-fn-seconds gets a wall-clock-dependent error string (the pass it was in), so near-budget functions can still differ run to run.