The minimal-diff reactive text edit deleted the differing span first
and inserted the replacement after. Deleting collapsed an
insertion-type-nil marker sitting on the FIRST character of the
preserved suffix onto the edit start, so it ended up stranded before
the inserted text instead of on its (unchanged) character -
contradicting the documented "markers sitting in unchanged text keep
their positions". The drift happened even for same-length
replacements ("0" -> "9").
tp--edit-region-minimal-diff now inserts the replacement at
edit-start first and then deletes the shifted old span
((delete-region (point) (+ (point) (- edit-end edit-start)))): the
insertion shifts the suffix-boundary marker right with its character
and the deletion pulls it back into place. Markers whose characters
were deleted now end at the END of the edit rather than the start
(documented side effect, still inside the replacement span).
Point clamping semantics are preserved: the documented "point inside
the edited span ends up at the start of the edit" was previously an
artifact of save-excursion's own marker collapsing under
delete-then-insert, which the new edit order breaks.
tp--replace-reactive-text-in-buffer therefore owns point restoration
now - it remembers point in a marker, clamps it to the edit start
when an edit swallowed it, and restores it at the end - and
tp--update-reactive-text's save-excursion wrapper is removed (it
would have overridden the clamp with its drifted marker). All four
shipped point/marker behavior tests pass unchanged.
New regressions port verify-txt1: suffix-start marker tracking for
grow, same-length and shrink edits (direct calls), the deleted-char
marker landing at the edit end, and a real setq-driven suffix-marker
case.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
tp--buffer-has-layer-region-p tested liveness with a direct tp-name
property search, which only sees the rendered top layer of each run.
A layer buried below a pushed top, or hidden via tp-hide-layer, keeps
its tp-name inside the tp-layers storage property, so:
- tp-gc-anonymous-layers undefined layers a live buffer still held,
destroying their definitions and reactive deps; pop/show later
resurfaced text whose reactivity was permanently dead;
- tp-reactive-track-buffer (the documented remedy for the
string-insertion registration gap) never registered buried or
hidden layers either, so the gap-closer itself had the gap.
Fix:
- tp-reactive.el gains the shared stack-aware scan
tp-reactive--buffer-layer-names: a tp--map-intervals walk that
collects the direct tp-name of each run plus the tp-name of every
plist inside the run's tp-layers value (read as a plain list of
plists, so tp-reactive needs nothing from the stack module).
tp-reactive-track-buffer is rewritten on it.
- tp--buffer-has-layer-region-p is rewritten on the same scan and
moves - together with tp-gc-anonymous-layers (autoload kept) - from
tp-render.el into tp-layer.el beside tp--anonymous-layer-registry,
so the whole anonymous-layer lifecycle (mint, intern, undefine,
collect) lives in one module and tp-render sheds its only
non-rendering responsibility. tp-render's tp--map-layer-buffers
keeps calling the predicate downward.
Regression tests port the review's gc1/xm02 probes: buried-under-push
kept alive with reactivity surviving a later pop; hidden layer kept
alive with show+setq re-rendering; track-buffer registering buried
and all-hidden layers from inserted strings. Existing tests cover
the visible-kept and killed-buffer-collected control cases.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The reactive engine located regions solely via the direct tp-name
text property - the render cache of the topmost visible layer - and
wrote only direct properties. The resolved-value snapshots stored
inside the tp-layers property (hidden layers, layers buried below the
top, and in full-stack mode even the visible top's own snapshot) were
never refreshed, so:
- an update received while a layer was hidden rendered stale values
after tp-show-layer, permanently (re-setting the same value is a
watcher no-op and could not repair it);
- with any layer hidden, the very next stack operation - even a no-op
tp-move-layer - rebuilt from the stale snapshot and silently
reverted the reactive update;
- a below-top reactive layer revealed by tp-pop-layer rendered its
stale snapshot;
- one buffer could render two different values of one variable at
once (mixed visible/hidden regions);
- reactive tp-text never reached hidden regions at all.
Fix in tp-render.el:
- tp--write-layer-through-stack-storage: for every run whose
tp-layers holds an entry of the updated layer, replace the layer's
own keys in that entry (preserving its tp-hidden flag) and rebuild
the run via the tp--stack-props-to-list / tp--stack-build-props
codec, which also refreshes the topmost-visible render cache in
full-stack mode. Unchanged runs are left untouched.
- tp--update-layer-regions calls it after the direct render pass.
- tp--replace-reactive-text-in-buffer gains a second pass replacing
text in regions where the layer sits only inside storage (text is
physical - hide/show toggles properties, never text), carrying all
existing properties over the edit, then writes the refreshed props
including tp-text through to storage. The minimal-diff edit is
factored into tp--edit-region-minimal-diff, shared by both passes.
Acceptance tests port the review's XM-01 probe scenarios: update
while hidden renders after show (A3); no-op stack ops never revert
and same-value re-set is never needed (B1-B4); pop reveals current
values (C2); reactive tp-text reaches hidden text (T1); mixed
visible/hidden regions stay in sync (X1); plus the hid1-probe
show+hide round-trip of an unrelated hidden top.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Reactive performance (tp-reactive/tp-render/tp-ops): layer->buffer
registry replaces full buffer-list scans on reactive updates, with a
conservative unknown->learning-scan fallback, kill-buffer pruning,
and tp-reactive-track-buffer to close the propertized-string-insert
gap; minimal-diff tp-text replacement preserves point and markers and
makes identical-text updates true no-ops; tp-gc-anonymous-layers
collects interned anonymous layers no registered buffer still shows.
Stack capabilities (tp-stack): tp-hide-layer/tp-show-layer visibility
(hidden layers stay in the stack but do not render; hiding the top
reveals the next visible layer), tp-lower-layer, tp-rotate-layer
DIRECTION/COUNT, tp-layer-stack-at, and modified-run-count returns
with NOERROR options.
Search and layer APIs (tp-search/tp-layer): SUBEXP capture groups and
START/END bounds for tp-regexp-*/tp-match-*, PREDICATE/NOT-CURRENT
exposure on tp-forward/tp-backward/-do (defaults keep 0.2.0 semantics),
multi-argument parameterized layers (define-tp/define-tps arglists of
any arity, tp-layer-props-with-args, tp-layer-arglist), and the
interactive tp-describe-layer.
79 new regression tests; combined suite 522/522 green, shuffled run
green, doctests 63/63, byte-compile clean with warnings-as-errors.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Byte-compile warnings swept 57 -> 0 across all modules, tests, and
doctest (docstring rewraps and quoting, defvar declarations for the
reactive test variables, prefixed doctest counters, dead-binding
removal, one impossible eq -> equal in a face-merge assertion) with
behavior preserved. GitHub Actions workflow runs an Emacs 28.1/29.4/
30.1 matrix: compile-all with warnings-as-errors, the 443-test suite,
a genuinely shuffled-order rerun (tp-run-shuffled.el runs each test
individually; ERT's member selector cannot reorder), and the 63
README doctests. Makefile gains WERROR, compile-all, and
test-shuffled. Autoload cookies added for the four interactive
commands and the define-tp/define-tps macros. package-lint: 0
findings (main file tp.el); draft MELPA recipe in docs/.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
35 fixes across tp-layer.el, tp-stack.el, tp-search.el, tp-reactive.el
and tp-render.el, each empirically reproduced before and after:
layer-definition resolution/cycles/copying/cleanup, clipped region-local
stack mutators, symmetric backward search matching, length-changing
replacements, and reactive re-render correctness (replace-not-accumulate,
buffer-local isolation, batching union, per-interval props).
Adds four per-module regression suites (100 new tests); combined suite
is 438/438 green. tp-test-backward updated to the now-symmetric backward
matching contract it previously codified inverted. The string-vs-buffer
shortfall divergence in tp-forward-do/tp-backward-do is documented in
CHANGELOG as a known divergence rather than changed. Adds CHANGELOG.md.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
tp.el (4866 lines) is now an umbrella over nine modules with an
enforceable dependency order: tp-core -> tp-reactive -> tp-layer ->
tp-ops -> tp-search -> tp-render -> tp-stack -> tp-palette ->
tp-builtins. Upward dependencies are inverted through four hook
variables installed by tp-render.el.
Also: require text-property-search (fixes tp-backward void-function),
clip tp-intervals to the requested range, add the shared clipping
interval walker tp--map-intervals and tp-face-properties, remove
synced-conflict junk files, byte-compilation now succeeds (define-tp
macroexpansion previously failed at compile time).
All 280 legacy tests pass plus 8 new tp-core tests.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
