9MjGSrSSKJr SSKrSSKJr SSKJrJrJ r SSK J r SSK J r SSKJr SS KJrJr SS KJr SS KJrJrJrJr SS KJr S rSSjrSSjrSSjrg)ucDetect and repair Camelot capture drops before structure correction. Camelot occasionally drops printed values during cell assignment even though its text extraction captured them: a totals row typeset on a raised baseline splits into two row bands at the default grouping tolerance, and the numbers' vertical center lands exactly on the band boundary, failing the strict containment test on both sides — the row's label and dollar signs survive, its numbers vanish from the grid. The repair loop keeps authority deterministic and puts the model only where judgement is needed: 1. DETECT (deterministic): numeric tokens present in the table region's text layer but absent from the grid. Same numeric normalization as the correction-stage grounding guard. 2. RECOMMEND (agent): the capture advisor sees the cropped table image, the extracted grid, and the dropped tokens with their printed lines, and proposes one bounded retry adjustment. 3. RETRY + ACCEPT (deterministic): re-run the region pass with the advised knobs. The retry replaces the original ONLY if it recovers every missing numeric and loses none the original had. Anything else keeps the original grid, so a wrong recommendation costs one extra Camelot pass and nothing else. Every detection and every verdict is logged — a capture drop on a financial document must never be silent. ) annotationsN)Path)ListOptionalSet)logger)CaptureAdvisor) page_words)CamelotCandidategrid_to_markdown) numeric_keys)WordBoxbbox_to_top_leftcrop_region_pngtable_crop_box)camelot_targetedg@c[SRSU555n[SRSU555nX#- $)zFNumeric tokens in the region's text layer that appear in no grid cell. c3*# UH oSv M g7f)N).0ws E/home/mande/repo/quber/src/quber/core/extractors/camelot/recapture.py 'dropped_numeric_keys..:s!= 1A$ sc36# UHoHo"v M M g7fNrrrowcs rrr;s AesSSe)r join)cells region_wordslayergrids rdropped_numeric_keysr(8s< != !== >E  Ae AA BD <c^/n[U5HinUHKm[TS5U1:XdMSRU4SjU55nURSUSU35 MR URSUS35 Mk U$)uoFor each missing token, the full printed line it sits on — the evidence the advisor grounds its diagnosis in.rrc3b># UH$n[USTS- 5S:dMUSv M& g7f)rN)abs)rxrs rr printed_lines..Fs0Slc!A$1+>NQR>R!ls/ /'z' on printed line: z' (word not isolated on a line))sortedr r#append)r%missinglineskeyliners @r printed_linesr8?sEgAAaD!cU*xxSlSS q%8?@  LL1SE!@A B Lr)c # UbURbUcU$[R"[[ U5U5IShvN upxn [ URU[ S9uppU Vs/sH0oSU :dM USU ::dMUSU :dM#USU ::dM.UPM2 nn[URU5nU(dU$[R"SU[U5[U55 [R"[U[URU5U5IShvN nURU[!UR5[#UU5U5IShvN nUcU$[R$"SUUR&UR(UR*UR,5 URunnnn[/UU5S S [1UU5S S [1UU5S S [/UU5S 3n[R"[2UUUUUR&UR(UR*5IShvN nUc[R"S U5 U$[7S R9SUR555n[7S R9SUR555nUU-nUU- nUU:Xa*U(d#[R$"SU[U55 U$[R"SU[U5[U5[U55 U$GNs snfGNGNN![4a$n[R"S UU5 UsSnA$SnAff=f7f)uReturn `cand`, or an advised retry that provably captures more. No-op when the advisor is off, the candidate has no bbox, or no drop is detected. The retry is scoped to the candidate's own bbox — the exact frame the drop was measured in. N)padrr,r-ujpage {}: capture drop detected — {} numeric token(s) in the table's text layer missing from the grid: {}zJpage {}: capture advisor recommends flavor={} row_tol={} column_tol={}: {}z.1f,z?page {}: capture-repair retry failed; keeping original grid: {}z=page {}: capture-repair retry found no grid; keeping originalrc36# UHoHo"v M M g7frrrs rr!repair_capture..s$J*3cQcQ*r"c36# UHoHo"v M M g7frrrs rrr>s#J#cAcAr"uEpage {}: capture repair ACCEPTED — recovered {} token(s), lost nonezRpage {}: capture repair REJECTED (recovered {}/{}, lost {}); keeping original grid)bboxasyncio to_threadr rrDETECT_PAD_PTSr(r$rwarninglenr2rr recommendr r8infoflavorrow_tol column_tol diagnosisminmaxr Exceptionr r#)candsourcepage page_imagedpiordinaladvisor_pwpage_hwordslefttoprightbottomrr%r4cropadvicex1y1x2y2arearetryexcoriginalretried recoveredlosts rrepair_capturerjNs$))+z/A &00T&\4PPC/ 6~VDu$ju!! A1A1Q4SV;A[\]^[_ci[iAuLj"4::|ML'M( L8LLLLBML" 8ML%BM0AL*2L(3L*7C$MM%M(L** M4M MMMM)r$zList[List[str]]r% List[WordBox]returnSet[str])r%rkr4rmrlz List[str])rOr rPstrrQintrRzOptional[Path]rSrorTrorUzOptional[CaptureAdvisor]rlr ) __doc__ __future__rrApathlibrtypingrrrlogururquber.agents.capture_advisorr quber.agents.completenessr %quber.core.extractors.camelot.acquirer r 7quber.core.extractors.camelot.correspondence.correctionr 5quber.core.extractors.camelot.correspondence.geometryrrrr5quber.core.extractors.camelot.correspondence.recoveryrrCr(r8rjrr)rr{s8#&&70TP S  T T T T T T  T&TTr)