diff --git a/pytest.ini b/pytest.ini
index 2ede9fb3e..9814b3365 100644
--- a/pytest.ini
+++ b/pytest.ini
@@ -10,3 +10,4 @@ addopts = --capture=no
 norecursedirs = tests/manual tests/evals scripts/cdp-download-poc tests/sdk eval infra/gcp-agent prompt_evaluation internal-tools
 markers =
     workday_offline: opt-in offline Workday deterministic regression suite (needs a Playwright chromium; run with -m workday_offline)
+    boundary: reads skyvern-frontend/ or docs/ from the repo, so it must run on FE/docs-only PRs that skip the Python suite
diff --git a/skyvern/forge/taskv3/loop.py b/skyvern/forge/taskv3/loop.py
index 70156ee71..55f8cccf3 100644
--- a/skyvern/forge/taskv3/loop.py
+++ b/skyvern/forge/taskv3/loop.py
@@ -203,7 +203,20 @@ PERCEPTION_STALL_REASON_PREFIX = "perception_stall:"
 # returning different content, or a download landing; a first-time probe has no baseline and is
 # evidence of nothing, so varied-selector probing cannot launder repetition into "progress".
 ACTION_LOOP_NUDGE_AFTER = 3
-ACTION_LOOP_TERMINATE_AFTER = 6
+# 8, not 6: RAW repeats of one action key peak at 6 across 50 completed prod runs while stuck runs
+# reach 36, so 6 sat on the completed population's edge. DO NOT LOWER THIS. The effective
+# post-clearing counter — what this constant is actually compared against — was then measured by
+# replaying the clearing rule over per-call telemetry, and NO SEPARATION WAS OBSERVED: two completed
+# runs peaked at 2 and 4, four stuck runs at 1, 3, 3 and below (upper bounds, computed identically
+# on both sides), and the highest observed value fell in a COMPLETED run. n=6 is far too small to
+# establish inversion as a property of either population — that spread is also consistent with
+# noise — but it is no evidence FOR a lower threshold either, and lowering a safety threshold
+# requires positive evidence. 8 is above everything observed, which is why the change is safe; it
+# is also why nothing in the sample fires. REVISIT if the effective distribution ever becomes
+# measurable at population scale; do not lower it before then. Repeat count may not be a stuck-ness
+# signal at all: SKY-15602 tracks the real problem, which is that the loop has no definition of
+# goal progress, only of page change.
+ACTION_LOOP_TERMINATE_AFTER = 8
 
 # Facetable sibling of PERCEPTION_STALL_REASON_PREFIX; same dashboard contract.
 ACTION_LOOP_REASON_PREFIX = "action_loop:"
@@ -1706,14 +1719,22 @@ async def run_agent_tool_loop(
             # (e.g. "next") on every page relies on THIS clear to survive — page_transitioned alone
             # deliberately does not clear the action-loop guard (see below), so only a progressed
             # snapshot does.
-            _clear_action_state()
+            ring = ledger.content_only.get(action_key) if content_only_digest is not None else None
+            # INVARIANT: this test and the budget-evidence test below must both refuse a return to
+            # the ring. They answer one question — did the run reach ground it has not already
+            # covered? — and relaxing either alone silently reopens SKY-14998: a page that CYCLES
+            # moves on every probe, so `progressed` holds every round, and an unconditional clear
+            # let the action DRIVING the oscillation reset its own counter forever (one production
+            # key ran 11 times against a threshold of 6 while the nudge fired once).
+            returned_to_known_ground = ring is not None and content_only_digest in ring
+            if not returned_to_known_ground:
+                _clear_action_state()
             # Two landed digests that differ are positive evidence, exactly like a fingerprint
             # mismatch — and the only movement evidence there is when page_fingerprint is absent.
             canonical.progress(_ProgressEvidence.PERCEPTION_DIGEST)
             # Evidence requires NEW content: a URL-only flip (history.pushState) still clears the
             # repeat guards above but earns no budget, and neither does a return to a content state
             # in the probe's recent ring (a panel toggling open and shut).
-            ring = ledger.content_only.get(action_key) if content_only_digest is not None else None
             if ring and content_only_digest not in ring:
                 _note_page_change_evidence()
         if content_only_digest is not None:
diff --git a/skyvern/forge/taskv3/tools.py b/skyvern/forge/taskv3/tools.py
index a8b992e88..b78fdb53f 100644
--- a/skyvern/forge/taskv3/tools.py
+++ b/skyvern/forge/taskv3/tools.py
@@ -760,14 +760,29 @@ def _lone_duplicate_candidate(rows: list[dict[str, Any]]) -> int | None:
     present_vals = {str(o.get("val")) for o in rows if o.get("val") is not None}
     if len(present_vals) >= 2:
         return None
-    # Sets must be NESTED, not equal: an a11y copy often mirrors only a subset of the primary row's
-    # attributes ({"55", "New York"} beside {"55"} is one candidate). Two sets neither containing the
-    # other carry a genuine disagreement and refuse.
-    val_sets = [frozenset(str(v) for v in o.get("vals") or []) for o in rows]
-    non_empty_sets = [s for s in val_sets if s]
-    for i, a in enumerate(non_empty_sets):
-        for b in non_empty_sets[i + 1 :]:
-            if not (a <= b or b <= a):
+    # Two-rule agreement over the surface+attribute-keyed entries. (1) A key present on BOTH rows
+    # must carry one value — crossed values across surfaces refuse, while a value carried on one
+    # row's surface only says nothing against the other row. (2) The depth-blind floor: one row's
+    # flattened attr=value pairs must nest inside the other's. Rows declaring identity on disjoint
+    # attributes cannot be confirmed the same, and an agreeing generic attribute (a shared title)
+    # beside disjoint identifiers is ordinary markup, not identity evidence — the pairs don't nest,
+    # so it refuses. A row declaring NOTHING still constrains nothing (its empty set nests, so an
+    # attribute-less a11y copy collapses).
+    val_maps: list[dict[str, str]] = []
+    for o in rows:
+        entries: dict[str, str] = {}
+        for raw in o.get("vals") or []:
+            key, _, val_part = str(raw).partition("=")
+            entries[key] = val_part
+        val_maps.append(entries)
+    for i, first in enumerate(val_maps):
+        for second in val_maps[i + 1 :]:
+            for shared in first.keys() & second.keys():
+                if first[shared] != second[shared]:
+                    return None
+            flat_first = {f"{k.partition(':')[2]}={v}" for k, v in first.items()}
+            flat_second = {f"{k.partition(':')[2]}={v}" for k, v in second.items()}
+            if not (flat_first <= flat_second or flat_second <= flat_first):
                 return None
     n = rows[0].get("n")
     return n if isinstance(n, int) else None
@@ -873,21 +888,43 @@ def _ambiguous_rows_error(
     )
 
 
-def _row_value_suffix(o: dict[str, Any]) -> str:
+def _row_value_suffix(o: dict[str, Any], rows: list[dict[str, Any]]) -> str:
     """Every present distinguishing surface for one row — value, accessible label, then any other
     declared value (data-code, data-key, ...) — each truncated to 60 chars like the row text already is,
     so a page-controlled attribute can never blow a refusal message up wholesale. Surfaces are additive
     (not first-match), since the veto may have come from a surface other than the first one present.
+    `rows` is the same-text row set the refusal lists: a surface that agrees across every row
+    distinguishes nothing, so the row-label clause prints only when the names disagree, and the
+    capped vals render entries that differ across rows first.
     """
     parts = ""
     if o.get("val") is not None:
         parts += f" (value {str(o.get('val'))[:60]!r})"
     if o.get("label"):
         parts += f" (label {str(o.get('label'))[:60]!r})"
+    # The ancestor's name is its own surface: when it differs from the preferred display label (a
+    # shared leaf label masking distinct row names) AND disagrees across the rows, the disagreeing
+    # name is the one that converts — gated like the veto itself, on cross-row disagreement.
+    ancestor_name = (o.get("labels") or [None, None])[1]
+    ancestor_names = {str((p.get("labels") or [None, None])[1]) for p in rows if (p.get("labels") or [None, None])[1]}
+    if ancestor_name and str(ancestor_name) != str(o.get("label") or "") and len(ancestor_names) >= 2:
+        parts += f" (row label {str(ancestor_name)[:60]!r})"
     # `vals` entries are attribute-keyed for comparison; render only the value part, and subtract
     # what the value surface already showed so a row never prints one value twice.
     shown_already = {str(o.get("val")).strip()} if o.get("val") is not None else set()
-    vals = [bare[:60] for v in (o.get("vals") or []) if (bare := str(v).split("=", 1)[-1]) not in shown_already]
+    raw_vals = [str(v) for v in (o.get("vals") or [])]
+    # The caller is being asked to pick between these rows: entries every row carries identically
+    # cannot be what tells them apart, so the ones that differ print first (the cap must not hide
+    # the distinguishing value behind agreeing generic ones).
+    common_to_all = set(raw_vals)
+    for p in rows:
+        if p is not o:
+            common_to_all &= {str(v) for v in (p.get("vals") or [])}
+    vals = [
+        bare[:60]
+        for v in sorted(raw_vals, key=lambda v: v in common_to_all)
+        if (bare := v.split("=", 1)[-1]) not in shown_already
+    ]
     if vals:
         shown_vals = vals[:3]
         more = "; ..." if len(vals) > 3 else ""
@@ -911,7 +948,7 @@ def _identical_text_rows_error(
         # and the stale tags may re-land on different rows at the next scan.
         return f'[data-tv3-sugg="{o.get("n")}"] ' if tags_live else ""
 
-    listing = "; ".join(f"{_sel(o)}{str(o.get('text') or '')[:60]!r}{_row_value_suffix(o)}" for o in shown)
+    listing = "; ".join(f"{_sel(o)}{str(o.get('text') or '')[:60]!r}{_row_value_suffix(o, rows)}" for o in shown)
     more = len(rows) - len(shown)
     next_step = (
         'click the intended row directly by its [data-tv3-sugg="N"] selector; the typed query was left '
@@ -4325,13 +4362,17 @@ _MENU_OPTION_TEXTS_JS = (
       // The DOM `.value` IDL is the spec submission value ALREADY -- an explicit `value=""` and an
       // absent attribute (which falls back to the option's own text) are genuinely distinct submission
       // values even though both can display identical text, so this always reads it, never null.
+      // NOT trimmed: the DOM preserves whitespace in option submission values, so "x" and "x " are
+      // genuinely distinct — the trim rule covers attribute-authoring drift, not submission values.
       val = String(el.value);
     } else {
       // Same presence rule as the OPTION branch: an attribute that EXISTS is a present value even
       // when empty -- `data-value=""` next to `data-value="x"` is a real disagreement, not absence.
       const dv = el.getAttribute('data-value');
       const va = el.getAttribute('value');
-      val = dv !== null ? dv : va;
+      // data-value is authoring metadata (whitespace drift between copies is presentational); a
+      // `value` attribute is a submission value and stays byte-exact like the OPTION branch above.
+      val = dv !== null ? String(dv).trim() : va !== null ? String(va) : null;
     }
     // The collapse's OWN value read: the verifier's declaredValues drops all-digit and >40-char
     // values on purpose (they must not CONFIRM a commit), but for telling two same-text rows apart a
@@ -4344,6 +4385,11 @@ _MENU_OPTION_TEXTS_JS = (
     const vetoVals = [];
     try {
       for (const node of new Set([el, opt || el])) {
+        // Keyed by SURFACE and attribute ("a:" the option row, "l:" a leaf inside it): crossed
+        // values across surfaces are disagreements a flat set cannot see. The key must come from
+        // DOM structure, not tagging depth — a self-tagged row and its twin tagged at a leaf must
+        // read the row's attributes under the same key, or equal renders refuse each other.
+        const where = node === (opt || el) ? 'a:' : 'l:';
         for (const a of node.attributes) {
           if (!VETO_ATTR.test(a.name)) continue;
           // Trimmed for comparison: whitespace drift between a portal copy and an inline copy is
@@ -4354,9 +4400,7 @@ _MENU_OPTION_TEXTS_JS = (
           // veto blind to distinct long values, while the fingerprint still tells apart any pair
           // differing in prefix or length. Only identical-prefix-same-length pairs read as equal.
           if (v.length > 512) v = v.slice(0, 512) + '#len' + v.length;
-          // Keyed by attribute: crossed values across surfaces (data-code="A" data-key="B" beside
-          // data-code="B" data-key="A") are a disagreement an unkeyed set cannot see.
-          vetoVals.push(a.name + '=' + v);
+          vetoVals.push(where + a.name + '=' + v);
         }
       }
     } catch (e) { /* attributes unreadable: no veto values */ }
@@ -4367,7 +4411,8 @@ _MENU_OPTION_TEXTS_JS = (
       setsize: Number.isFinite(setsize) && setsize > 0 ? setsize : 0,
       pos: pos,
       val: val,
-      vals: vetoVals.slice(0, 12),
+      // 9 allowlisted attributes x 2 nodes = 18 possible entries; 24 can never truncate.
+      vals: vetoVals.slice(0, 24),
       // Names are FULL veto inputs (truncation happens only where a message renders them): a slice
       // here would read two labels diverging past the cut as equal and collapse distinct rows.
       label: (accessibleName(el) || accessibleName(opt)) || null,
@@ -5180,6 +5225,9 @@ async () => {
   // diverge whenever a retained control is dropped later for having no selector that names it.
   let hiddenKept = 0;
   let hiddenListed = 0;
+  // Candidates the visibility gates below drop. Those drops are silent, so a page whose whole app
+  // shell is behind a boot gate renders exactly like an empty one; this is what tells the two apart.
+  let hiddenDropped = 0;
   let phantomDropped = 0;
   let truncated = 0;
   let truncatedInComponents = 0;
@@ -5293,12 +5341,12 @@ async () => {
     // like v1 (whose center_x check is only reached for a non-zero rect), so the zero-size
     // skinned-proxy carve-out below still runs for an off-screen-positioned skinned control.
     const centerX = (gr.left + gr.width) / 2 + window.scrollX;
-    if (ownGated && gr.width !== 0 && gr.height !== 0 && centerX < 0 && !_hScrolledAncestor(gateEl)) { continue; }
+    if (ownGated && gr.width !== 0 && gr.height !== 0 && centerX < 0 && !_hScrolledAncestor(gateEl)) { hiddenDropped++; continue; }
     // v1's isElementStyleVisibilityVisible (domUtils.js) drops a control whose own computed
     // visibility is not 'visible'. Scoped to non-zero-rect elements so the zero-size skinned-proxy
     // carve-out below still runs; visibility is read per-element, so a visibility:visible child of a
     // hidden ancestor is kept. A native checkbox/radio judges the parent here instead of itself.
-    if (ownGated && gr.width !== 0 && gr.height !== 0 && window.getComputedStyle(gateEl).visibility !== 'visible') { continue; }
+    if (ownGated && gr.width !== 0 && gr.height !== 0 && window.getComputedStyle(gateEl).visibility !== 'visible') { hiddenDropped++; continue; }
     let hidden = false;
     if (r.width === 0 || r.height === 0) {
       // Design systems skin a native SELECT/checkbox/radio/file input at zero size behind a styled
@@ -5319,6 +5367,7 @@ async () => {
         // only when it actually renders visible content, as v1's recursion does -- an empty,
         // all-hidden, or all-off-canvas host is a phantom.
       } else {
+        hiddenDropped++;
         continue;
       }
     }
@@ -5834,7 +5883,7 @@ async () => {
     }
   } catch (e) { iframeInfo.total = 0; iframeInfo.inComponents = 0; iframeInfo.entries.length = 0; iframeInfo.unread = 0; iframeInfo.failed = true; }
 
-  return JSON.stringify({ url: location.href, title: document.title, text: texts, textFull: texts.map((t) => { const f = fullText.get(t); return f && f !== t ? f : null; }), textTruncated: textFull, textDropped: textDropped, iframes: iframeInfo, dropped: dropped, truncated: truncated, truncatedInComponents: truncatedInComponents, unnamedAnonymous: unnamedAnonymous, unnamedBudget: unnamedBudget, unnamedDuplicated: unnamedDuplicated, unnamedUnverifiable: unnamedUnverifiable, unnamedUnsafe: unnamedUnsafe, unreadableRoot: sawUnreadableRoot, undiscoveredRoots: undiscoveredRoots, rootCount: allRoots.length - 1, hiddenListed: hiddenListed, phantomDropped: phantomDropped, markersMinted: markersWritten, markersReused: markersReused, pageMutated: mutated, elements: out });
+  return JSON.stringify({ url: location.href, title: document.title, text: texts, textFull: texts.map((t) => { const f = fullText.get(t); return f && f !== t ? f : null; }), textTruncated: textFull, textDropped: textDropped, iframes: iframeInfo, dropped: dropped, truncated: truncated, truncatedInComponents: truncatedInComponents, unnamedAnonymous: unnamedAnonymous, unnamedBudget: unnamedBudget, unnamedDuplicated: unnamedDuplicated, unnamedUnverifiable: unnamedUnverifiable, unnamedUnsafe: unnamedUnsafe, unreadableRoot: sawUnreadableRoot, undiscoveredRoots: undiscoveredRoots, rootCount: allRoots.length - 1, hiddenListed: hiddenListed, hiddenDropped: hiddenDropped, phantomDropped: phantomDropped, markersMinted: markersWritten, markersReused: markersReused, pageMutated: mutated, elements: out });
 }
 """
 )
@@ -6577,6 +6626,20 @@ def build_browser_tools(
             lines.append(
                 f"note: {hidden_kept} native control(s) hidden behind styled proxies are listed with [hidden-native]"
             )
+        hidden_dropped = data.get("hiddenDropped") or 0
+        # Scoped to total blindness: present-but-hidden chrome (closed menus, inactive tabs) is on
+        # nearly every page, so an unconditional note would cost the prefix on every call to say
+        # nothing. With no element listed the count is the whole signal -- it separates an app shell
+        # still behind its boot gate from a page that genuinely has no controls, which is the one
+        # distinction the model cannot otherwise make and will poll wait->observe for turns to guess.
+        # Deliberately time-neutral: a boot gate resolves on its own and a stuck one never does, and
+        # this cannot tell which, so the wording must not imply that waiting is what fixes it.
+        if not elements and hidden_dropped:
+            lines.append(
+                f"note: the page has {hidden_dropped} control(s) that are present but not visible "
+                "(CSS-hidden or positioned off-canvas): the DOM is populated but none of it is "
+                "currently actionable"
+            )
         phantom_dropped = data.get("phantomDropped") or 0
         if phantom_dropped:
             lines.append(
@@ -6753,6 +6816,7 @@ def build_browser_tools(
         summary = {
             "text_dropped": text_dropped,
             "hidden_listed": hidden_kept,
+            "hidden_dropped": hidden_dropped,
             "phantom_dropped": phantom_dropped,
             "iframes_in_component_roots": iframe_info.get("inComponents") or 0,
             "undiscovered_roots": data.get("undiscoveredRoots") or 0,
@@ -8753,7 +8817,8 @@ def build_browser_tools(
             def _listed_row(o: dict[str, Any], canon_text: str) -> str:
                 entry = f'[data-tv3-menu="{o.get("n")}"] {str(o.get("text") or "")[:60]!r}'
                 if text_counts[canon_text] >= 2:
-                    entry += _row_value_suffix(o)
+                    same_text = [p for p, pt in zip(shown, shown_canon_texts) if pt == canon_text]
+                    entry += _row_value_suffix(o, same_text)
                 return entry
 
             listing = "; ".join(_listed_row(o, t) for o, t in zip(shown, shown_canon_texts))
diff --git a/tests/unit/test_api_docs_taxonomy.py b/tests/unit/test_api_docs_taxonomy.py
index d834ca44a..ddab6a23c 100644
--- a/tests/unit/test_api_docs_taxonomy.py
+++ b/tests/unit/test_api_docs_taxonomy.py
@@ -15,6 +15,10 @@ sync workflow, and `scripts/sync_openapi_docs.py --check`.
 import json
 from pathlib import Path
 
+import pytest
+
+pytestmark = pytest.mark.boundary
+
 COMMITTED_SPEC = Path(__file__).resolve().parents[2] / "docs" / "api-reference" / "openapi.json"
 
 # Capitalized resource tags — kept in lockstep with .agents/skills/api-docs-audit/SKILL.md.
diff --git a/tests/unit/test_docs_webhook_verification_snippets.py b/tests/unit/test_docs_webhook_verification_snippets.py
index afd6fd31b..dda995831 100644
--- a/tests/unit/test_docs_webhook_verification_snippets.py
+++ b/tests/unit/test_docs_webhook_verification_snippets.py
@@ -18,6 +18,8 @@ import pytest
 
 from skyvern.forge.sdk.core.security import generate_skyvern_webhook_signature
 
+pytestmark = pytest.mark.boundary
+
 REPO_ROOT = Path(__file__).resolve().parents[2]
 SNIPPET = REPO_ROOT / "docs/snippets/webhook-signature-verification.mdx"
 
diff --git a/tests/unit/test_taskv3_loop.py b/tests/unit/test_taskv3_loop.py
index c6048c8e9..6d56bb963 100644
--- a/tests/unit/test_taskv3_loop.py
+++ b/tests/unit/test_taskv3_loop.py
@@ -28,7 +28,9 @@ from skyvern.forge.taskv3 import loop as loop_module
 from skyvern.forge.taskv3.loop import (
     ACTION_BUDGET_EXTENDED_EVENT,
     ACTION_BUDGET_EXTENSION_REFUSED_EVENT,
+    ACTION_LOOP_NUDGE_AFTER,
     ACTION_LOOP_REASON_PREFIX,
+    ACTION_LOOP_TERMINATE_AFTER,
     CANONICAL_SURVIVAL_EVENT,
     FAILURE_EVIDENCE_MIN_TOOL_CALLS,
     FAILURE_EVIDENCE_MIN_TURNS,
@@ -3543,6 +3545,44 @@ async def test_action_loop_catches_varied_probe_evasion() -> None:
     assert caller.calls <= 15
 
 
+def test_action_loop_terminate_threshold_is_pinned_to_its_measured_value() -> None:
+    # The other action-loop tests read this constant so they track policy instead of drifting, which
+    # leaves nothing asserting the VALUE — someone could set it to 50 and the suite would stay green.
+    # Pinning it makes any change deliberate and visible in a diff, and sends the reader to the
+    # do-not-lower note at the constant: the effective post-clearing counter was measured, showed no
+    # separation between completed and stuck runs, and its highest observed value fell in a completed
+    # run — so there is no positive evidence supporting a lower threshold.
+    assert ACTION_LOOP_TERMINATE_AFTER == 8
+    assert ACTION_LOOP_NUDGE_AFTER < ACTION_LOOP_TERMINATE_AFTER
+
+
+@pytest.mark.asyncio
+async def test_action_loop_survives_a_page_that_oscillates_between_two_known_states() -> None:
+    # The production shape the guard was structurally blind to (SKY-14998, tsk in wr_568475173904014164):
+    # one action key ran 11 times against a terminate threshold of 6, and the repeat nudge fired
+    # exactly ONCE at repeat_count=3 before the run died on the token cap. A page that CYCLES rather
+    # than freezes moves on every probe, so `snap.progressed` held every round and wiped the whole
+    # repeat ledger — the action driving the oscillation reset its own counter forever. Returning to
+    # a state this probe has already seen is not progress and must not clear the guard.
+    from skyvern.forge.taskv3.loop import ACTION_LOOP_REASON_PREFIX
+
+    panel = ["url=x text: 'filters panel open'", "url=x text: 'filters panel shut'"]
+    contents = [panel[i % 2] for i in range(16)]
+    script: list[list[tuple[str, dict[str, Any]]]] = []
+    for _ in range(16):
+        script.append([("click", {"selector": "#apply"})])
+        script.append([("observe", {})])
+    clicks: list[tuple[str, dict[str, Any]]] = []
+    tools = [_billable_tool("click", clicks), _perception_tool("observe", contents), make_finish_tool()]
+    outcome, _ = await _run(script, tools, max_turns=200, max_tool_calls=500)
+
+    assert outcome.status == "terminated"
+    assert outcome.reason.startswith(ACTION_LOOP_REASON_PREFIX)
+    assert "#apply" in outcome.reason
+    # Bounded well below the 16 the script offers, and below the 11 the production run reached.
+    assert len(clicks) <= 10, len(clicks)
+
+
 @pytest.mark.asyncio
 async def test_pagination_with_changing_page_content_never_trips_action_loop() -> None:
     # Healthy pagination clicks the same Next selector many times, but each page's observe differs —
@@ -3637,7 +3677,8 @@ async def test_action_loop_warn_and_terminate_emit_facetable_logs() -> None:
     warned = [entry for entry in logs if entry["event"] == "taskv3 loop action repeat nudged"]
     terminated = [entry for entry in logs if entry["event"] == "taskv3 loop action repeated"]
     assert len(warned) == 1 and warned[0]["tool"] == "click" and warned[0]["repeat_count"] == 3
-    assert len(terminated) == 1 and terminated[0]["tool"] == "click" and terminated[0]["repeat_count"] == 6
+    assert len(terminated) == 1 and terminated[0]["tool"] == "click"
+    assert terminated[0]["repeat_count"] == ACTION_LOOP_TERMINATE_AFTER
 
 
 @pytest.mark.asyncio
@@ -4071,7 +4112,9 @@ async def test_warn_always_precedes_terminate_even_after_single_batch_burst() ->
     from skyvern.forge.taskv3.loop import ACTION_LOOP_REASON_PREFIX
 
     script = [
-        [("click", {"selector": "#submit"})] * 5,
+        # The burst must cross the terminate threshold inside ONE turn, or the property under test
+        # (no verdict before a delivered warning) is never exercised.
+        [("click", {"selector": "#submit"})] * ACTION_LOOP_TERMINATE_AFTER,
         [("click", {"selector": "#submit"})],
         [("click", {"selector": "#submit"})],
     ]
@@ -4080,7 +4123,8 @@ async def test_warn_always_precedes_terminate_even_after_single_batch_burst() ->
     outcome, _ = await _run(script, tools, max_turns=200, max_tool_calls=500)
     assert outcome.status == "terminated"
     assert outcome.reason.startswith(ACTION_LOOP_REASON_PREFIX)
-    assert len(clicks) == 7  # 5 burst + 1 post-warn-queue + 1 post-warn-delivery
+    # burst + 1 post-warn-queue + 1 post-warn-delivery
+    assert len(clicks) == ACTION_LOOP_TERMINATE_AFTER + 2
     warns = [m for m in outcome.messages if m.get("role") == "user" and "#submit" in str(m.get("content"))]
     assert len(warns) == 1
     assert outcome.messages.index(warns[0]) < len(outcome.messages) - 1
@@ -4095,11 +4139,11 @@ async def test_action_loop_counts_errored_attempts() -> None:
     clicks: list[tuple[str, dict[str, Any]]] = []
     click = _recording_tool("click", clicks, raises=True)
     click.billable = True
-    script = [[("click", {"selector": "#dead"})] for _ in range(10)]
+    script = [[("click", {"selector": "#dead"})] for _ in range(ACTION_LOOP_TERMINATE_AFTER + 4)]
     outcome, _ = await _run(script, [click, make_finish_tool()], max_turns=200, max_tool_calls=500)
     assert outcome.status == "terminated"
     assert outcome.reason.startswith(ACTION_LOOP_REASON_PREFIX)
-    assert len(clicks) == 6
+    assert len(clicks) == ACTION_LOOP_TERMINATE_AFTER
 
 
 @pytest.mark.asyncio
diff --git a/tests/unit/test_taskv3_tools.py b/tests/unit/test_taskv3_tools.py
index 3b43d6f06..a38e229f9 100644
--- a/tests/unit/test_taskv3_tools.py
+++ b/tests/unit/test_taskv3_tools.py
@@ -1679,6 +1679,7 @@ async def test_observe_result_carries_count_only_summary_for_the_call_record() -
     assert set(summary) == {
         "text_dropped",
         "hidden_listed",
+        "hidden_dropped",
         "phantom_dropped",
         "iframes_in_component_roots",
         "undiscovered_roots",
@@ -16226,12 +16227,15 @@ def _button_listbox_html(
     anchor_id: str = "cc",
     labels: list[str | None] | None = None,
     labelledby: list[str | None] | None = None,
-    wrap_in_span: bool = False,
+    wrap_in_span: bool | list[bool] = False,
     span_label: str | None = None,
+    span_attrs: list[dict[str, str]] | None = None,
     extra_attrs: list[dict[str, str]] | None = None,
 ) -> str:
     labels = labels or [None] * len(rows)
     extra_attrs = extra_attrs or [{} for _ in rows]
+    span_attrs = span_attrs or [{} for _ in rows]
+    wraps = wrap_in_span if isinstance(wrap_in_span, list) else [wrap_in_span] * len(rows)
     # `labelledby[i]` is the TEXT of a hidden div this row's aria-labelledby points at (not an id) --
     # the id is minted from the row's own position so each row gets a distinct label target.
     labelledby = labelledby or [None] * len(rows)
@@ -16248,7 +16252,9 @@ def _button_listbox_html(
         for k, v in extra_attrs[i].items():
             attrs += f" {k}={v!r}"
         span_attr = f" aria-label={span_label!r}" if span_label else ""
-        inner = f'<span style="cursor:pointer"{span_attr}>{text}</span>' if wrap_in_span else text
+        for k, v in span_attrs[i].items():
+            span_attr += f" {k}={v!r}"
+        inner = f'<span style="cursor:pointer"{span_attr}>{text}</span>' if wraps[i] else text
         return f'<li role="option"{attrs}>{inner}</li>'
 
     li_html = "".join(
@@ -16778,6 +16784,95 @@ async def test_select_combobox_refuses_rows_with_crossed_values_across_surfaces(
         assert r.status == "error", r.content
 
 
+def test_lone_duplicate_candidate_refuses_crossed_leaf_and_ancestor_values() -> None:
+    # The node dimension matters: leaf/ancestor data-code A/B beside A/A share the leaf value but
+    # disagree on the ancestor's — a flat set would read the second as a subset and collapse.
+    from skyvern.forge.taskv3.tools import _lone_duplicate_candidate
+
+    rows = [
+        {"n": 1, "text": "Depot", "vals": ["l:data-code=A", "a:data-code=B"]},
+        {"n": 2, "text": "Depot", "vals": ["l:data-code=A", "a:data-code=A"]},
+    ]
+    assert _lone_duplicate_candidate(rows) is None
+
+
+def test_lone_duplicate_candidate_collapses_same_value_at_a_different_depth() -> None:
+    # A copy carrying the same value on a different node is structural variance, not disagreement:
+    # only keys present on BOTH rows compare.
+    from skyvern.forge.taskv3.tools import _lone_duplicate_candidate
+
+    rows = [
+        {"n": 1, "text": "Depot", "vals": ["l:data-code=55"]},
+        {"n": 2, "text": "Depot", "vals": ["a:data-code=55"]},
+    ]
+    assert _lone_duplicate_candidate(rows) == 1
+
+
+@_skip_no_browser
+@pytest.mark.asyncio
+async def test_select_combobox_commits_duplicate_rows_with_whitespace_drifted_data_value() -> None:
+    # The primary value surface trims like every other veto surface: data-value "x " beside "x" is
+    # whitespace drift between copies, not a disagreement.
+    rows: list[tuple[str, str | None]] = [("Depot", "x "), ("Depot", "x")]
+    async with _content_page(_duplicate_suggestion_html(rows)) as page:
+        tools = build_browser_tools(_fixed_page_provider(page))
+        r = await _tool(tools, "select_combobox").handler({"selector": "#addr", "value": "Depot"})
+        assert r.status == "ok", r.content
+        value = await page.eval_on_selector("#addr", "el => el.value")
+        assert value == "Depot", value
+
+
+@_skip_no_browser
+@pytest.mark.asyncio
+async def test_shared_leaf_label_refusal_names_the_disagreeing_ancestor_labels() -> None:
+    # When a shared leaf action label masks distinct row names, the refusal must surface the names
+    # that actually disagree, or both rows print identically beside their selectors.
+    rows: list[tuple[str, str | None]] = [("Depot", None), ("Depot", None)]
+    async with _content_page(
+        _button_listbox_html(rows, labels=["Branch A", "Branch B"], wrap_in_span=True, span_label="Choose")
+    ) as page:
+        tools = build_browser_tools(_fixed_page_provider(page))
+        r = await _tool(tools, "select_combobox").handler({"selector": "#cc", "value": "Depot"})
+        assert r.status == "error", r.content
+        assert "Branch A" in r.content, r.content
+        assert "Branch B" in r.content, r.content
+
+
+@_skip_no_browser
+@pytest.mark.asyncio
+async def test_select_combobox_refuses_rows_declaring_identity_on_disjoint_attributes() -> None:
+    # Two rows that each declare identity on entirely different surfaces cannot be confirmed the
+    # same candidate (data-code "A" beside data-key "B" refused on main and must keep refusing).
+    rows: list[tuple[str, str | None]] = [("Depot", None), ("Depot", None)]
+    async with _content_page(_duplicate_suggestion_html(rows, attrs=[{"data-code": "A"}, {"data-key": "B"}])) as page:
+        tools = build_browser_tools(_fixed_page_provider(page))
+        r = await _tool(tools, "select_combobox").handler({"selector": "#addr", "value": "Depot"})
+        assert r.status == "error", r.content
+
+
+def test_lone_duplicate_candidate_refuses_same_attribute_with_disjoint_values_across_depths() -> None:
+    # The same surface naming disjoint values at different depths is a disagreement even though no
+    # node+attr key is shared.
+    from skyvern.forge.taskv3.tools import _lone_duplicate_candidate
+
+    rows = [
+        {"n": 1, "text": "Depot", "vals": ["l:data-code=A"]},
+        {"n": 2, "text": "Depot", "vals": ["a:data-code=B"]},
+    ]
+    assert _lone_duplicate_candidate(rows) is None
+
+
+@_skip_no_browser
+@pytest.mark.asyncio
+async def test_select_combobox_refuses_option_rows_with_whitespace_distinct_submission_values() -> None:
+    # The DOM preserves whitespace in option submission values: "x " and "x" are distinct choices.
+    rows: list[tuple[str, str | None]] = [("Foo", "x "), ("Foo", "x")]
+    async with _content_page(_duplicate_suggestion_option_html(rows)) as page:
+        tools = build_browser_tools(_fixed_page_provider(page))
+        r = await _tool(tools, "select_combobox").handler({"selector": "#addr", "value": "Foo"})
+        assert r.status == "error", r.content
+
+
 def _grid_suggestion_html(rows: list[tuple[str, str | None]]) -> str:
     # The ARIA grid-combobox pattern: the tagged leaf is the gridcell, and the candidate's identity
     # (data-code here) lives on its [role=row] ancestor.
@@ -16838,6 +16933,199 @@ async def test_select_combobox_commits_grid_duplicate_rows_sharing_the_row_code(
         tools = build_browser_tools(_fixed_page_provider(page))
         r = await _tool(tools, "select_combobox").handler({"selector": "#addr", "value": "Depot"})
         assert r.status == "ok", r.content
+
+
+@_skip_no_browser
+@pytest.mark.asyncio
+async def test_select_combobox_refuses_disjoint_identifiers_despite_shared_generic_attribute() -> None:
+    # A generic attribute agreeing across rows (a shared tooltip) is ordinary markup, not identity
+    # evidence: rows declaring identity on disjoint attributes must keep refusing even when a title
+    # happens to agree on both.
+    rows: list[tuple[str, str | None]] = [("Depot", None), ("Depot", None)]
+    async with _content_page(
+        _duplicate_suggestion_html(
+            rows, attrs=[{"data-code": "A", "title": "Choose"}, {"data-key": "B", "title": "Choose"}]
+        )
+    ) as page:
+        tools = build_browser_tools(_fixed_page_provider(page))
+        r = await _tool(tools, "select_combobox").handler({"selector": "#addr", "value": "Depot"})
+        assert r.status == "error", r.content
+        committed = await page.eval_on_selector("#addr", "el => el.getAttribute('data-committed')")
+        assert committed is None, committed
+
+
+def test_lone_duplicate_candidate_refuses_disjoint_identifiers_with_shared_generic_attribute() -> None:
+    # Any agreeing shared attribute must not stand in for the disjoint identifiers beside it: the
+    # flattened pair sets do not nest, so there is no affirmative evidence of one candidate.
+    from skyvern.forge.taskv3.tools import _lone_duplicate_candidate
+
+    rows = [
+        {"n": 1, "text": "Depot", "vals": ["a:data-code=A", "a:title=Choose"]},
+        {"n": 2, "text": "Depot", "vals": ["a:data-key=B", "a:title=Choose"]},
+    ]
+    assert _lone_duplicate_candidate(rows) is None
+
+
+def test_lone_duplicate_candidate_refuses_agreement_on_one_identifier_beside_disjoint_declarations() -> None:
+    # Deliberately conservative: rows agreeing on their one shared attribute while each declaring
+    # other, disjoint surfaces still refuse — a false collapse silently commits the wrong row, while
+    # a refusal only hands the pick back to the caller.
+    from skyvern.forge.taskv3.tools import _lone_duplicate_candidate
+
+    rows = [
+        {"n": 1, "text": "Depot", "vals": ["a:data-code=A", "a:title=T"]},
+        {"n": 2, "text": "Depot", "vals": ["a:data-code=A", "l:name=N"]},
+    ]
+    assert _lone_duplicate_candidate(rows) is None
+
+
+@_skip_no_browser
+@pytest.mark.asyncio
+async def test_select_combobox_commits_equal_renders_tagged_at_different_depths() -> None:
+    # One candidate rendered as a wrapped leaf beside a flattened copy: the tagger lands on the span
+    # in one row and on the row element itself in the other. Tagging depth is structural variance,
+    # not identity — the same DOM must read the same way and collapse.
+    rows: list[tuple[str, str | None]] = [("Depot", None), ("Depot", None)]
+    async with _content_page(
+        _button_listbox_html(
+            rows,
+            wrap_in_span=[True, False],
+            extra_attrs=[{"title": "Depot, Main St"}, {"title": "Depot, Main St"}],
+            span_attrs=[{"title": "Depot"}, {}],
+        )
+    ) as page:
+        tools = build_browser_tools(_fixed_page_provider(page))
+        r = await _tool(tools, "select_combobox").handler({"selector": "#cc", "value": "Depot"})
+        assert r.status == "ok", r.content
+        value = await page.eval_on_selector("#cc-value", "el => el.value")
+        assert value == "Depot", value
+
+
+@_skip_no_browser
+@pytest.mark.asyncio
+async def test_select_combobox_refuses_leaf_values_disagreeing_under_agreeing_row_values() -> None:
+    # The mirror of the crossed-depth case: leaves declare A beside B while both rows declare Z. A
+    # depth-blind read would let the agreeing row value shadow the leaf disagreement and collapse
+    # two distinct rows.
+    rows: list[tuple[str, str | None]] = [("Depot", None), ("Depot", None)]
+    async with _content_page(
+        _button_listbox_html(
+            rows,
+            wrap_in_span=True,
+            extra_attrs=[{"data-code": "Z"}, {"data-code": "Z"}],
+            span_attrs=[{"data-code": "A"}, {"data-code": "B"}],
+        )
+    ) as page:
+        tools = build_browser_tools(_fixed_page_provider(page))
+        r = await _tool(tools, "select_combobox").handler({"selector": "#cc", "value": "Depot"})
+        assert r.status == "error", r.content
+        value = await page.eval_on_selector("#cc-value", "el => el.value")
+        assert value == "", value
+
+
+@_skip_no_browser
+@pytest.mark.asyncio
+async def test_select_combobox_commits_duplicate_rows_carrying_the_same_attribute_at_both_depths() -> None:
+    # The accept direction of the two-depth read: twin renders agreeing at the leaf AND the row.
+    rows: list[tuple[str, str | None]] = [("Depot", None), ("Depot", None)]
+    async with _content_page(
+        _button_listbox_html(
+            rows,
+            wrap_in_span=True,
+            extra_attrs=[{"data-code": "Z"}, {"data-code": "Z"}],
+            span_attrs=[{"data-code": "A"}, {"data-code": "A"}],
+        )
+    ) as page:
+        tools = build_browser_tools(_fixed_page_provider(page))
+        r = await _tool(tools, "select_combobox").handler({"selector": "#cc", "value": "Depot"})
+        assert r.status == "ok", r.content
+        value = await page.eval_on_selector("#cc-value", "el => el.value")
+        assert value == "Depot", value
+
+
+@_skip_no_browser
+@pytest.mark.asyncio
+async def test_select_combobox_refuses_rows_whose_distinguishing_value_sits_past_twelve_entries() -> None:
+    # The value list must never truncate ahead of the disagreeing entry: nine allowlisted leaf
+    # attributes plus the row's own push the rows' disagreeing data-code to position 13, which a
+    # 12-entry cap discarded — both rows then read identical and the first was committed.
+    nine = {
+        "value": "v",
+        "data-value": "v",
+        "data-val": "v",
+        "data-v": "v",
+        "data-code": "C",
+        "data-key": "k",
+        "data-option-value": "o",
+        "name": "n",
+        "title": "t",
+    }
+    rows: list[tuple[str, str | None]] = [("Depot", None), ("Depot", None)]
+    async with _content_page(
+        _button_listbox_html(
+            rows,
+            wrap_in_span=True,
+            span_attrs=[dict(nine), dict(nine)],
+            extra_attrs=[
+                {"title": "t", "name": "n", "value": "v", "data-code": "ROW-A"},
+                {"title": "t", "name": "n", "value": "v", "data-code": "ROW-B"},
+            ],
+        )
+    ) as page:
+        tools = build_browser_tools(_fixed_page_provider(page))
+        r = await _tool(tools, "select_combobox").handler({"selector": "#cc", "value": "Depot"})
+        assert r.status == "error", r.content
+        value = await page.eval_on_selector("#cc-value", "el => el.value")
+        assert value == "", value
+
+
+@_skip_no_browser
+@pytest.mark.asyncio
+async def test_select_combobox_refuses_rows_with_whitespace_distinct_value_attribute() -> None:
+    # A `value` attribute is a submission value like OPTION.value, not authoring metadata: "x " and
+    # "x" are two choices, so byte-exact comparison covers it and only data-value gets the drift trim.
+    rows: list[tuple[str, str | None]] = [("Depot", None), ("Depot", None)]
+    async with _content_page(_duplicate_suggestion_html(rows, attrs=[{"value": "x "}, {"value": "x"}])) as page:
+        tools = build_browser_tools(_fixed_page_provider(page))
+        r = await _tool(tools, "select_combobox").handler({"selector": "#addr", "value": "Depot"})
+        assert r.status == "error", r.content
+
+
+@_skip_no_browser
+@pytest.mark.asyncio
+async def test_identical_text_refusal_omits_row_labels_that_agree_across_rows() -> None:
+    # The row-label clause exists to surface the surface that DISAGREES; when every row wears the
+    # same ancestor name the clause distinguishes nothing and must not print at all.
+    rows: list[tuple[str, str | None]] = [("Depot", "v1"), ("Depot", "v2")]
+    async with _content_page(
+        _button_listbox_html(rows, labels=["Same Branch", "Same Branch"], wrap_in_span=True, span_label="Choose")
+    ) as page:
+        tools = build_browser_tools(_fixed_page_provider(page))
+        r = await _tool(tools, "select_combobox").handler({"selector": "#cc", "value": "Depot"})
+        assert r.status == "error", r.content
+        assert "row label" not in r.content, r.content
+
+
+@_skip_no_browser
+@pytest.mark.asyncio
+async def test_identical_text_refusal_leads_with_the_value_that_disagrees() -> None:
+    # Agreeing generic values must not crowd the disagreeing identity out of the capped suffix —
+    # the caller is being asked to pick between the rows, so what differs prints first.
+    rows: list[tuple[str, str | None]] = [("Depot", None), ("Depot", None)]
+    async with _content_page(
+        _duplicate_suggestion_html(
+            rows,
+            attrs=[
+                {"title": "T", "name": "N", "data-val": "V", "data-code": "AAA"},
+                {"title": "T", "name": "N", "data-val": "V", "data-code": "BBB"},
+            ],
+        )
+    ) as page:
+        tools = build_browser_tools(_fixed_page_provider(page))
+        r = await _tool(tools, "select_combobox").handler({"selector": "#addr", "value": "Depot"})
+        assert r.status == "error", r.content
+        assert "AAA" in r.content, r.content
+        assert "BBB" in r.content, r.content
         value = await page.eval_on_selector("#addr", "el => el.value")
         assert value == "Depot", value
 
@@ -20849,3 +21137,58 @@ async def test_semantic_verify_kill_switch_keeps_commits_working(monkeypatch: py
             {"selector": "#city", "value": "Springfield, Sangamon, IL"}
         )
         assert picked.status == "ok", picked.content
+
+
+@_skip_no_browser
+@pytest.mark.asyncio
+async def test_observe_discloses_that_a_populated_page_was_dropped_whole_by_the_visibility_gates() -> None:
+    # An app shell that keeps its content in the DOM behind a boot gate renders `(0 interactive
+    # elements)` -- byte-identical to a genuinely empty page, because the visibility gates drop
+    # silently. That ambiguity is what makes a model poll wait->observe for turns on end: nothing in
+    # the payload says the page is full and unreadable rather than still loading. v1 is equally blind
+    # here (measured), so the gates stay; only the disclosure is new.
+    html = (
+        "<!doctype html><html><head><title>Portal</title>"
+        "<style>#shell { visibility: hidden; }</style></head><body>"
+        '<div id="shell">'
+        '<nav><a href="#/home">Home</a><a href="#/logout">Logout</a></nav>'
+        "<h1>Document Portal</h1><p>Welcome back.</p>"
+        '<button id="newdoc">New Document</button>'
+        '<input id="q" type="text" placeholder="Search documents">'
+        '<div role="status">Signed in successfully</div>'
+        "</div>"
+        # Outside the hidden shell, so it is the off-canvas gate that drops this one and not the
+        # visibility gate: the count then discriminates both gates rather than only the second.
+        '<button id="ghost" style="position:absolute;left:-9999px">Ghost</button>'
+        "</body></html>"
+    )
+    async with _content_page(html) as page:
+        tools = build_browser_tools(_fixed_page_provider(page))
+        r = await _tool(tools, "observe").handler({})
+
+    assert r.status == "ok", r.content
+    assert "(0 interactive elements)" in r.content, r.content
+    assert "note: the page has 5 control(s) that are present but not visible" in r.content, r.content
+    assert r.data is not None and r.data["summary"]["hidden_dropped"] == 5
+
+
+@_skip_no_browser
+@pytest.mark.asyncio
+async def test_observe_stays_silent_about_hidden_chrome_when_it_can_still_see_the_page() -> None:
+    # The note is scoped to total blindness on purpose: site chrome is routinely present-but-hidden
+    # (closed menus, inactive tabs), so a note on every observe would be noise on the one channel
+    # whose cost is ~linear in the prefix. A page the model can act on says nothing.
+    html = (
+        "<!doctype html><html><head><title>Portal</title></head><body>"
+        '<button id="go">Go</button>'
+        '<div style="display:none"><button id="hidden-menu">Archive</button></div>'
+        "</body></html>"
+    )
+    async with _content_page(html) as page:
+        tools = build_browser_tools(_fixed_page_provider(page))
+        r = await _tool(tools, "observe").handler({})
+
+    assert r.status == "ok", r.content
+    assert "(1 interactive elements)" in r.content, r.content
+    assert "present but not visible" not in r.content, r.content
+    assert r.data is not None and r.data["summary"]["hidden_dropped"] == 1