refactor(tui): one viewport representation, delete the line-index anchors

[zipadoodlez]

Epic #1411, phase 6a. Draft: tentatively stacked on #1477, #1526 and #1429.

The chat viewport kept the reader's position in four places at once: a stored
scroll_offset, an auto_scroll_paused flag, a distance-from-bottom
HistoryScrollAnchor for a prepend, and a PendingResizeAnchor for a rewrap.
Every pair of them had to be reconciled whenever the transcript moved, and the
guards that reconciled them were themselves a source of bugs.

This stores one thing instead:

follow: Option<ContentPos>,   // None = live tail, Some = anchored to content

ContentPos names an item id and a row, which survives a rewrap, a prepend and a
compaction because it is content, not a row index. The render path resolves it
against the frame it is drawing and stores nothing back.

Both anchors existed only to compensate for the stored row offset, so they go
with it:

• a resize needs no capture, because a content position is width-independent;
• a prepend needs no anchor, because the item id cannot move.

The upward overshoot that fed the prepend anchor is dropped with it (see the
ponytail: note in maybe_queue_compacted_history_load); the content anchor
already keeps the view steady across the load.

The motion telemetry goes too

anchor_stability.rs plus ui_smoothness.rs (~810 production lines) watched the
viewport for jarring motion: content repositioning away from its anchor,
insertions above, single-frame pops, blinks, mass reflows. Three reasons it
belongs in this phase and not a later one:

• Its input is the state this change deletes: each frame is fed
  app.scroll_offset() and !app.auto_scroll_paused(), i.e. exactly the two
  fields 6a removes.
• It cannot do its job. Phase 1 established that a rewrap changes every row hash,
  so a correct reflow and a jump are indistinguishable to it — the mass-reflow
  signal fires on precisely the case it was built to catch.
• Every phase since has been verified by a behavioural regression test instead,
  so what remains is a debug surface measuring a projection of state that no
  longer exists.

The smoothness and smoothness:reset debug commands, and
smoothness_benchmark.rs, go with it.

Supporting changes

• content_pos_at_row / resolve_content_pos now treat a section prepared
  without a wrap map (the padded header, inline images, batch progress) as one
  raw line per rendered row, so those rows can hold an anchor instead of being
  silently unmappable.
• message_row_ranges returns an iterator rather than allocating a Vec of
  every message boundary, because the position is now resolved every frame.

Why it is stacked. 6a deletes the state that #1477 (6b), #1526 (5c) and
#1429 (5b) each convert, so its base has to contain all three. Please land those
first. Until they do, this PR's diff carries their commits; it rebases to a
single commit on master once they land.

Size: 2512 deletions against 520 insertions, net -1992 (production -1084,
tests -908) in the single 6a commit.

Tests. resize_anchor.rs and smoothness_benchmark.rs tested the removed
machinery and are deleted. The invariant resize_anchor.rs covered now lives in
scroll_copy_03: a_prepend_keeps_the_reader_on_the_same_content and
a_resize_keeps_the_reader_on_the_same_content assert the anchored content stays
under the reader across a prepend and a rewrap (including a round trip back to
the original width).

The rest of the migration is the encoding change: the old scroll_offset meant
distance from the bottom while following the tail and an absolute row while
parked, so assertions like scroll_offset == 0 for "at the bottom" become
!follow.is_some(), and the up/down comparisons flip. Tests that parked the
viewport before the first render now draw once first, because a content anchor
needs a frame to name a row; those with no terminal set the follow target
directly (park_scroll).

Validation:

• cargo fmt --all --check clean.
• cargo clippy --profile selfdev -p jcode-tui-core -p jcode-tui-messages -p jcode-tui --all-targets -- -D warnings clean.
• cargo test -p jcode-tui --lib -- --test-threads=1: 2421 passed, 0 failed.
• cargo test -p jcode-tui-core: 48 passed. cargo test -p jcode-tui-messages: 18 passed.

Under the default parallel harness a handful of model-picker, session-restore and
info-widget tests fail. Those are pre-existing: the same tests fail the same way
on this branch's base (5f2e66adb), including a login test that dies on a DNS
error, and all of them pass single-threaded.

Deliberately kept, not oversights: scroll_offset() and auto_scroll_paused()
remain on TuiState as derived reads (34 call sites in metrics/debug/diff would
otherwise move), and last_total_wrapped_lines stays a published per-frame read
rather than being recomputed from the frame.