e4f5f778

By: ValentinHirschi <valentin.hirschi@gmail.com>

reuse exact CFF bases and compact completed products

Keep separate builder-local caches for plain lower-sector and direct CFF
bases, keyed by their complete inputs and preserving the distinct source
sign conversions. Reuse completed expressions before numerator embedding.
Fuse equal chains at completed lower-component and public root-sample
boundaries. Private child bases retain zero-map carriers for later append.
No numerator expansion, graph, physics setting or test expectation changes.

The captured twelve-edge, two-loop CFF input completed in the scratch
prototype in 617.75 seconds, with a 14.48 GB peak under the 30 GB process-tree
guard. Three existing fixtures produced byte-identical complete outputs
with the optimization disabled and enabled. This is diagnostic evidence,
not a complete GL16 or scalar-matrix result.

Independent cache/boundary source review, formatting and diff checks pass.
Strict all-eleven-package release checking passes in 32.15 s (peak 1.773 GB),
and literal CI Clippy passes in 25.56 s (peak 1.418 GB), both with normal guards.
The independent eval-enabled CFF gate passes 119/119 tests in 4.142 s, with
156.333 s guarded compile/run time and peak 5.175 GB. Fresh GL16/165, focused 55
and literal local full 1995 gates remain pending; the fresh GL16 run has started. The previous 7b candidate passed
standard Linux/macOS and Nix CI; matching new-head CI is still required.
This remains a draft validation candidate, not ALL_GREEN.

User prompt:
Yes, continue as planned, it seems like it should be ok memory wise, so just let it run until LU scalar matrix is complete, and then really finalise the ALL_GREEN PR as mentioned earlier.

Related user instruction:
Ok continue as planned now, trying to finish this damn scalar LU matrix, and only looking for improvement to the code if you can find relatively non-invasive cheap wins. The core objective is getting to ALL_GREEN commit and push merge-ready state of the corresponding PR. Then stop yourself and deem the goal achieved (if there does not seem to be any gross oversight w.r.t what is in it).

Total

42m04s

Time to Start Worker time Duration Time to finish Idle
Config 1s 9s 9s 11s 1s
Eval 7s 47s 47s 54s 0s
Build 48s 1h53m21s 39m03s 39m51s 6s
Test 2m28s 24m10s 39m33s 42m01s 0s
Deploy 42m03s 2s 2s 42m06s 2s
Suite 1s 2h18m32s 42m04s 42m06s 10s

Timeline

Config Eval Build Test Deploy Compressed gap
0s1m2m3m4m5m6m7m8m9m10m11m12m13m14m15m16m17m19m20m21m22m23m24m25m26m27m28m29m30m31m32m33m34m35m36m37m38m39m40m41m42m