Move outstanding work from TODO.md into the issue tracker

TODO.md carried two records in one file: what had been done, with the
measurements behind it, and what was left. The second half is what a tracker
is for, and keeping it here has already cost something -- AGENTS.md records a
round where eleven entries described code that had already changed, and this
file admitted to three more.

Every open item is now an issue on source.starfort.tech/andrew/libakgl,
labelled by kind and blast radius and milestoned by what it can land in: 0.9.x
for anything that breaks no ABI, 0.10.0 for new or changed public symbols,
1.0.0 for the design work. Four are epics: the performance plan (#60),
coverage (#61), actor rotation (#62), and the false header comments (#63).

Verified against the tree before filing rather than transcribed. Three entries
were already fixed and were not filed: the akgl_path_relative context leak, the
akgl_draw_background test extension, and the SDL enumeration audit -- keyboards,
gamepads and mappings are all freed in CLEANUP today. Two were reworded because
the code had moved: the fonts item is a missing teardown entry point rather than
a missing API, since akgl_text_unloadallfonts exists, and draw_world's tilemap
call is already bounded by numlayers, so only the per-layer actor rescan remains.

TODO.md keeps the part a tracker has no place for: why a decision went the way
it did, what the measurement was, and which arguments turned out to be wrong.

TODO.txt is deleted. Four of its eight entries had shipped -- actor-to-actor
collision, actor-to-world collision, automatic facing, image layers -- and the
four that had not are #74 through #77, with the GPU renderer's research links
kept because that is the part that took the time.

Every reference that named an item number or a moved section is repointed, in
the manual, the headers, the tests and the examples.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

docs/17-text-and-fonts.md

@@ -127,8 +127,8 @@ Measuring is 340 times cheaper than drawing, which tells you the whole cost is t
 and the upload. **Six HUD readouts is 76 µs a frame. That is fine at 60 fps on a laptop and
 it is 4% of a 2 ms GPU frame** — and it is being paid every frame for a score that changes
 once a second. A one-line cache keyed on (font, string, colour) would take it to nothing;
-`PERFORMANCE.md` calls it the single clearest optimisation in the library, and `TODO.md`
-carries it as a target.
+`PERFORMANCE.md` calls it the single clearest optimisation in the library, and it is
+issue #22.
 
 So: **fine for a HUD line, wrong for a static body of text redrawn every frame.** If you are
 drawing a page of dialogue that does not change, rasterize it yourself once with SDL3_ttf,
@@ -209,7 +209,7 @@ guards, so drawing nothing still refuses everything drawing something refuses.
 > refused, and `text.h:120-122` still warns that a failure after rasterizing leaks the
 > surface and the texture. Both describe pre-0.5.0 behaviour. `src/text.c:110-112` returns
 > success for `""`, and `src/text.c:140-146` destroys both objects in a `CLEANUP` block that
-> runs on every path. `TODO.md` items 27 and 23 record both as fixed. The header comments
+> runs on every path. `TODO.md` records both as fixed. The header comments
 > want correcting in their own commit.
 
 ## Fonts, the pools, and what is not shared