docs: one ranked TODO.md, and a home for the ld370/as370 reference material

[mgrossmann]

Replaces two local working notes (TODO-LD370.md, TODO-ASM370.md — never committed, so nothing is lost from history) with one ranked TODO.md plus the reference material they carried.

Why the split

The notes mixed three things: open defects, settled history, and reference material about IFOX00 and the linkage editor. Only the first belongs in a tracker, only the third belongs in docs/, and the second is mostly noise once it is fixed. So:

• Live defects → issues. ld370: a WX parsed before an ER leaves a hard unresolved reference unreported — rc 0, zero adcon, S0C4 at runtime #99–as370: an unrecognised command-line option is silently taken as the source filename #104, six of them, each with a reproducer and a code citation. ld370: a WX parsed before an ER leaves a hard unresolved reference unreported — rc 0, zero adcon, S0C4 at runtime #99 (a WX parsed before an ER leaves a hard unresolved reference unreported — rc 0, zero adcon) was measured on the host during the review, not just derived; the order dependence reproduces in four lines of assembler.
• Reference material → docs/ld370-iewl-divergences.md and as370/docs/ifox-option-parity.md. Both cite routines and line numbers rather than reproducing IBM source; the provenance note explains why that line is drawn where it is.
• Already fixed → dropped. Most of the as370 note was closed by as370: DC/DS types P Z E L S Q assemble to zero bytes with RC 0, shifting every later symbol #53 (the DC/DS silent-zero class and the float paths) and by CXD becoming a proper "not implemented" diagnostic.

TODO.md

Modelled on mvslovers/httpd's: the tracker stays the source of truth, and this file holds what it cannot — the order, the reason for it, and what waits on a decision. The ranking rule is the project's own: silent wrong output before silent under-reporting before a loud gap before cosmetics, and inside the top class, having already shipped broken code breaks the tie. That puts #39 (MNOTE severity swallowed — the FREEMAIN it silently deleted in libc370 leaked exactly when storage was short) first.

Two corrections the review turned up

• Feature request: add "(BIND on <date> <time> V03 M08)" section to load modules to allow ISPF to show CHANGED date #13 was closeable and is now closed. PR Add BIND Time/Date to IDR #46 — from the reporter himself — implemented the BIND-date IDR in August and nobody closed the issue. Checked against the IEWL oracle member: we emit two IDR records where IEWL emits three. The remainder is the 21-byte translator IDR, and as370 already writes the data into every END card, so it is a propagation gap, not a new mechanism.
• cc370 silently drops -Wl,--ac,N, and ld370 --pack drops the AC again: an unauthorized module that looks authorized #37 is half solved, and the open half is not what it says. build_userdata() has carried a packed member's complete PDS2 user-data — AC, RENT/REUS included — since June, for -iebcopy inputs. The issue packs a bare .lm, which has no directory to carry anything from. Its "adjacent observation" (--norent byte-identical) is not a bug either: attributes live only in the directory entry. Both recorded in the issue.

Also here

• docs/tool-roadmap.md — the roadmap for tools beyond the four that build (objdump370, nm370, iebcopy370, …), rewritten in English, with libobj370/libmvs370 as the recommended Phase 0.
• docs/entry-point-resolution.md — cited by Startup customization via same-named @@START is fragile under autocall (multi-module) #10 and by the ld370 notes, but existing only locally until now. Facts re-checked against the tree before committing.

Documentation only; no code changes.

[mgrossmann]

Added a second commit: docs/roadmap-integration.md rewritten in English against the current state. The June version listed multi-track transport, the mbt host backend, the one-command driver and the built-in C library as open — all four landed since — and it was the last German document in the repository. It now points at TODO.md for the open work instead of carrying a list of its own, which is how it went stale in the first place.