qvm: candidate bound restoration

docs/ASSET_PIPELINE_TODO.md

@@ -565,3 +565,6 @@ Research progress: RE-627 [report](research/QVM_TRACE_RETRY.md); remaining nativ
 
 
 Research progress: RE-628 [report](research/QVM_TRACE_RELINK.md); remaining native/content/runtime/platform scope stays open.
+
+
+Research progress: RE-629 [report](research/QVM_TRACE_BOUNDS_RESTORE.md); remaining native/content/runtime/platform scope stays open.

docs/FINDINGS.md

@@ -917,3 +917,6 @@ Bounded qagame paths recover bomb-plant/inactivity correction requests, direct f
 
 
 **RE-628 — Skipped target relink ordering and duplicate entries:** The skipped-target cleanup issues native−31 in insertion order and preserves duplicate entries; empty lists issue no relink request. [Evidence](research/QVM_TRACE_RELINK.md).
+
+
+**RE-629 — Candidate bound restoration and f32 round-trip limits:** The restoration suffix repeats eligibility gates, reverses bound arithmetic and writes fixed ±15 dimensions rather than restoring a saved record. [Evidence](research/QVM_TRACE_BOUNDS_RESTORE.md).

docs/PROVISIONAL_TODO.md

@@ -885,3 +885,6 @@ Research progress: RE-627 [report](research/QVM_TRACE_RETRY.md); remaining nativ
 
 
 Research progress: RE-628 [report](research/QVM_TRACE_RELINK.md); remaining native/content/runtime/platform scope stays open.
+
+
+Research progress: RE-629 [report](research/QVM_TRACE_BOUNDS_RESTORE.md); remaining native/content/runtime/platform scope stays open.

docs/RESEARCH_LEDGER.md

@@ -817,3 +817,8 @@ Each retry captures unlink−32, appends the target pointer and increments count
 ## 2026-10-20 — RE-628 Skipped target relink ordering and duplicate entries (separate task commit)
 
 The skipped-target cleanup issues native−31 in insertion order and preserves duplicate entries; empty lists issue no relink request. Eight empty/reordered/duplicate/ten-entry cases independently verify pointer order, loop cursor and untouched list contents. Full suite: 997 tests pass. Report regenerates byte-identically (24991 bytes). Storage/filter, links and whitespace checks pass. No assets/Rust runtime/shared VM evaluator changed; native/reference/legal/full platform scope remains open.
+
+
+## 2026-10-20 — RE-629 Candidate bound restoration and f32 round-trip limits (separate task commit)
+
+The restoration suffix repeats eligibility gates, reverses bound arithmetic and writes fixed ±15 dimensions rather than restoring a saved record. Six canonical/arbitrary/large-magnitude fixtures independently compare all six candidate records and demonstrate exact and lossy round trips. Full suite: 1001 tests pass. Report regenerates byte-identically (647300 bytes). Storage/filter, links and whitespace checks pass. No assets/Rust runtime/shared VM evaluator changed; native/reference/legal/full platform scope remains open.

docs/REVERSE_ENGINEERING_TODO.md

@@ -928,3 +928,6 @@ Purpose: turn the supplied installation into an evidence-backed specification of
 
 
 - [x] RE-628 Skipped target relink ordering and duplicate entries ([report](research/QVM_TRACE_RELINK.md)).
+
+
+- [x] RE-629 Candidate bound restoration and f32 round-trip limits ([report](research/QVM_TRACE_BOUNDS_RESTORE.md)).

docs/research/QVM_TRACE_BOUNDS_RESTORE.md

@@ -0,0 +1,7 @@
+# Candidate bound restoration and f32 round-trip limits
+
+RE-629. Reproduce with `python3 scripts/analyze_qvm_trace_bounds_restore.py`; check with `python3 scripts/test_qvm_trace_bounds_restore.py`. The [machine report](qvm-trace-bounds-restore.json) retains source identity and bounded fixtures.
+
+Actual292326–292503 revisits the candidate index list and the same self/client3096/4324/3740 gates. It subtracts the widening constant from476/480/484, adds it to464/468, subtracts15 from456 and sets448/452=15 and436/440=−15. Fields472/444 remain untouched. This is arithmetic plus canonical-dimension writes, not a snapshot copy. The full-record oracle verifies all220 words for each candidate. Ordinary canonical dimensions round-trip exactly in the low-magnitude fixture; arbitrary prior dimensions and some large f32 values do not. Mid-trace state changes could alter restoration eligibility and remain unresolved until real trace/damage flow is recovered. No native effects or live-world/legal-input parity is claimed.
+
+No original executable/full QVM/native collision/physics/rendering executes. Native responses and legal input domains are supplied or unverified where stated; full runtime/macOS/Linux Rust/Bevy/Solari gameplay remains open.

docs/research/README.md

@@ -486,3 +486,6 @@ The extension collector adds five focused tests for native tail calls, command-r
 
 
 [RE-628: Skipped target relink ordering and duplicate entries](QVM_TRACE_RELINK.md). Reproduce with `python3 scripts/analyze_qvm_trace_relink.py`; check with `python3 scripts/test_qvm_trace_relink.py`.
+
+
+[RE-629: Candidate bound restoration and f32 round-trip limits](QVM_TRACE_BOUNDS_RESTORE.md). Reproduce with `python3 scripts/analyze_qvm_trace_bounds_restore.py`; check with `python3 scripts/test_qvm_trace_bounds_restore.py`.

scripts/analyze_qvm_trace_bounds_restore.py

@@ -0,0 +1,28 @@
+#!/usr/bin/env python3
+"""Actual candidate-bound restoration after the previously recovered widening."""
+import itertools
+from pathlib import Path
+from qvm_research_support import source,execute
+from analyze_qvm_shot_trace_request import fixture,first_trace,LOCAL
+from analyze_qvm_entity_init import POOL,STRIDE
+from analyze_qvm_temp_event import record
+from analyze_qvm_acceleration import word
+from analyze_qvm import write_json
+
+def restore(ins,m):
+    points=[POOL+m[34992592+4*j]*STRIDE for j in range(m[34992580])];before={str(p):record(m,p) for p in points};m,_,e=execute(ins,292326,292504,m,(0,0,0),LOCAL)
+    return m,{'initial':before,'final':{str(p):record(m,p) for p in points},'execution':e}
+
+def main():
+    q,raw,h,ins=source();rows=[]
+    for canonical,magnitude in itertools.product([False,True],[1.25,16777216.0,1e9]):
+        m,a=fixture(raw,h)
+        for index in [65,69]:
+            p=POOL+index*STRIDE
+            for o in [464,468,476,480,484,456]:m[p+o]=word(magnitude)
+            if canonical:
+                for o in [448,452]:m[p+o]=word(15)
+                for o in [436,440]:m[p+o]=word(-15)
+        original={str(POOL+i*STRIDE):record(m,POOL+i*STRIDE) for i in range(64,70)};m,pre=first_trace(ins,m,a);m,r=restore(ins,m);r.update({'original':original,'canonical_fixture':canonical,'magnitude':magnitude,'prelude_execution':pre['prelude'],'exact_original_restored':r['final']==original});rows.append(r)
+    write_json(Path('docs/research/qvm-trace-bounds-restore.json'),{'source':q['source'],'sha256':q['sha256'],'probes':rows,'limits':'Actual widening291642 prefix followed by292326–292503 restoration under unchanged candidate/client gates. Native requests captured. Restoration uses fixed dimensions and f32 inverse arithmetic, not saved-record restoration. Mid-trace candidate/state mutations and legal world domains remain unverified.'})
+if __name__=='__main__':main()

scripts/test_qvm_trace_bounds_restore.py

@@ -0,0 +1,29 @@
+import json,unittest
+from pathlib import Path
+from test_qvm_ground import word,n,f
+
+def restored(r):
+    r=dict(r)
+    for o in [476,480,484]:r[str(o)]=word(f(n(r[str(o)])-n(1103626240)))
+    for o in [464,468]:r[str(o)]=word(f(n(r[str(o)])+n(1103626240)))
+    for o in [448,452]:r[str(o)]=1097859072
+    r['456']=word(f(n(r['456'])-n(1097859072)))
+    for o in [436,440]:r[str(o)]=(-1049624576)&0xffffffff
+    return r
+class TraceBoundsRestoreTests(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):cls.r=json.loads(Path('docs/research/qvm-trace-bounds-restore.json').read_text())
+    def test_independent_full_record_oracle(self):
+        for p in self.r['probes']:
+            for address,r in p['initial'].items():self.assertEqual(p['final'][address],restored(r) if (int(address)-34091368)//880 in [65,69] else r)
+    def test_canonical_low_magnitude_round_trip(self):
+        p=next(p for p in self.r['probes'] if p['canonical_fixture'] and p['magnitude']==1.25);self.assertTrue(p['exact_original_restored'])
+    def test_arbitrary_extents_are_not_snapshots(self):
+        for p in self.r['probes']:
+            if not p['canonical_fixture']:self.assertFalse(p['exact_original_restored'])
+        self.assertTrue(any(not p['exact_original_restored'] for p in self.r['probes'] if p['canonical_fixture']))
+    def test_no_native_or_extra_field_restoration(self):
+        for p in self.r['probes']:
+            self.assertEqual(p['execution']['calls'],[])
+            for address,r in p['final'].items():self.assertEqual(r['472'],p['original'][address]['472']);self.assertEqual(r['444'],p['original'][address]['444'])
+if __name__=='__main__':unittest.main()