A lit shelf for the training-failure literature, with both papers filed

[syncytium2]

Files the two papers that #679 used to withdraw the chorus-collapse page's novelty
claim. They were read, acted on and cited publicly — and filed nowhere, which is the
exact failure lit/README.md exists to prevent:

A PDF with no entry is indistinguishable from a PDF someone downloaded and forgot.

New shelf rather than a stretched one. DL/ is learned detectors and set-structured
inputs; ml/ is the unlabelled detector's claims. Neither covers why training fails in
its first few hundred steps, which is what these two are about and what goal 2 will keep
running into.

<darkroom>/bugarach/lit/optimization/ now holds:

file
kosson_2024_warmup_update_size.pdf	Kosson, Messmer, Jaggi — NeurIPS 2024, arXiv:2410.23922, 22 pp
lu_2020_dying_relu_initialization.pdf	Lu, Shin, Su, Karniadakis — Commun. Comput. Phys. 2020, doi:10.4208/cicp.OA-2020-0165, 34 pp
README.md	both entries, in the shelf's author / year / provenance / which-decision form

Plus one row in lit/README.md's subfolder table. Nothing else under lit/ touched —
the shelf is new, so it cannot collide with the ACTIVE lit/radar/ claim.

The entries carry the distinction, not just the citation, so a later session re-reading
them does not have to redo #679's ruling: Kosson's units are alive and the update kills them
(counted at end of training, remedy a leaky ReLU — itself the evidence the mechanism is
ReLU's exact zero region); Lu's are dead before training and his theorem sends them to a
constant function, with a depth-up / width-down ordering that matches Table 1 in all four
comparisons at lr 0.03. Neither transfers to a GELU, so the README says plainly that the
ordering match is a match in direction, not evidence that his mechanism operates here.

It also records what the papers do not settle — including the one search nobody has run:
whether a learned event detector collapsing to a single call covering a whole recording has
been reported in the imaging literature. That is the remaining way the page's narrowed claim
could turn out to be old.

Verification: both PDFs fetched by hand from arXiv (fetch_paper.py is deliberately not
vendored — SAP004), and each checked after the copy by extracting its first page, not by file
size. Venues confirmed independently: Kosson from the arXiv comment field ("Accepted to
NeurIPS 2024"), Lu from its DOI.

One correction carried forward: role 2's archived report cites Lu's constant-function result
as Theorem 3.4; in the text it is 3.7. Nothing shipped wrong — the page carries no theorem
number — but the archive does.

🤖 Generated with Claude Code