…batch

Add trust-scored template matching (ambiguity / PSR)

edge_lines.find_grid recovers a bordered table's geometry but leaves the cells
empty; OCR gives the text but no structure, and nothing joined them. Drop OCR
boxes into the grid (assigned by cell-centre, gated by an overlap fraction),
concatenate each cell's text in reading order, flag merged-cell spans, and
convert to records / CSV. Pure-stdlib over plain dicts.

…ill-batch

Add table_grid_fill: fill a ruling-line grid with OCR text

serialize_observation renders one full frame (blows the token budget every
turn); element_diff gives the stable-ID correspondence but stops at element
pairs. Add the missing serializer: diff two frames, classify matched elements
as changed or stable, render only the churn as +/~/- lines (added & changed
first, stable dropped, capped at max_lines). Reuses element_diff.match_elements
and observation.observation_index.

…delta-batch

Add token-budgeted observation delta (what changed)

Every match_template_all call forces a hand-tuned min_score: too low floods
NMS, too high drops re-themed targets, and the right value differs per asset.
Run Otsu on the correlation score histogram to find the valley between
background correlation and real matches, returning that cut-off plus a
separability score so a unimodal (no-match) surface is flagged. match_auto
returns one peak per above-threshold region via connected_boxes.

ccoeff_normed scores span [-1, 1], so the Otsu cut-off can legitimately be
negative on some OpenCV builds (match_auto clamps it with floor anyway). Assert
the threshold is below a perfect match and use a relative separability check
(bimodal > flat) instead of absolute bounds.

…resh-batch

Add auto-thresholded template matching (Otsu on score map)

ocr/structure detects a table only when every row's cell-left-x matches, so it
fails on ragged / borderless / right-aligned columns; edge_lines.find_grid
needs ruling lines a whitespace table has none of. Find columns by the gaps:
project OCR boxes onto the x-axis, read the persistent empty vertical bands as
gutters, assign column indices, bucket rows by spacing, emit the table. Pure
difference-array projection, no numpy.

…t-batch

Add column_layout: infer columns from whitespace (borderless tables)

…tate

ocr/structure only pairs a label: with the immediately next cell, so it can't
handle label-above-value, two-column key/value, right-aligned values or
non-text widgets, and has no checkbox notion. Pair each label with the nearest
aligned value across directions (right/below) within max_gap, match
free-standing widgets to their nearest label, and read checkbox state from the
box's dark-pixel fill ratio.

…batch

Add form_fields: multi-direction label/value association + checkbox state

screen_state/element_diff report what changed but never tie it to the action;
loop_guard only flags a no-op after the same digest repeats N times. Diff the
before/after observation and, given the action's target point, classify the
result on the first step as no_op / changed_near_target / changed_elsewhere /
changed, returning the changed centres and a reason. Reuses
element_diff.match_elements and observation_delta's field-change check.

…t-batch

Add action_effect: classify did-my-click-do-anything with attribution

Intensity NCC drops when a control is re-filled or re-themed, and ORB needs
corner texture flat-design glyphs lack. Match by edge shape instead: Canny both
images, distance-transform the scene edges, slide the template's edges over it
and score by mean edge-to-edge distance (Chamfer). A perfect outline aligns at
~0 cost regardless of fill. Reuses visual_match's loaders/resize/NMS/Match and
edge_lines's Canny default.

…atch

Add edge_match: edge-shape (Chamfer) template matching

expect_poll/assert_eventually poll a single condition with no action-bound spec
and no before-baseline, so they can't express 'a new dialog appeared';
trajectory_eval is whole-trajectory. Evaluate a small JSON spec of clauses
(appears/disappears/enabled/disabled/text_present/text_absent/count) against
the after-observation, diffed against the before-observation, returning a
per-clause pass/fail report. compile_postcondition yields an after->bool
predicate for expect_poll.

…n-batch

Add postcondition: declarative expected-outcome specs for actions

self_healing/locator_repair only fix a locator that didn't resolve; loop_guard
only detects a stuck loop with no tactic selection or backoff. Consume an
effect verdict and drive a bounded retry loop, choosing the next untried tactic
(wait_retry/relocate/nudge/scroll_into_view/escalate) each round. Pure-stdlib
state machine with injected act/verify/apply_tactic/verdict_for/sleep seams.

…batch

Add step_repair: repair-tactic policy for failed / no-effect actions