I'll start by locating and profiling the input files. Let me first check the skill's reference materials and then examine the target folder. Ran 2 commands Found 83 training plan .xlsx files. Let me read the skill references and profile a few spreadsheets to understand their structure. Read 2 files, ran a command openpyxl and pandas are available; cluster.py uses only stdlib difflib, so the missing rapidfuzz is fine. Let me profile the spreadsheets — including checking for embedded hyperlinks, since "video URLs" are likely hyperlinks on exercise-name cells. Ran Profile first 4 spreadsheets Console encoding issue. Let me write profiling to a script file with UTF-8 output. Created a file, ran a command, read a file Now I understand the structure clearly. These are training plans where: Column A = exercise names (plus day headers Dzień A1/B1, rozgrzewka warmups, header rows) Columns B–E = Tempo / Serie / Powtórzenia / Przerwa (tempo/sets/reps/rest) Columns F+ = weekly loads Each file has multiple date-named sheets (training blocks), plus some empty ArkuszN sheets Video URLs are embedded inline in the exercise text (e.g. ...wiosłowanie Meadowsa https://mikolaj.onlinecourse..., https://youtu.be/...) Exercise names carry a superset prefix (A , B , C ) and athlete-specific cues after * Let me build the extraction and also check for URLs stored as real cell hyperlinks (not just inline text). Created a file, ran a command, read a file Rich results. Key findings: 11,509 exercise rows → 2,950 exact-unique base names (heavy variant sprawl — lots of semantic dedup needed). 2,606 cell hyperlinks; 527 unique names carry a video URL. The dominant source is the coach's own course platform mikolaj.onlinecoursehost.com/courses/.../lessons/N, plus some YouTube/Instagram. The lesson URL is a strong grouping signal: many name variants point to the same lesson (e.g. lesson /23 = the whole "biakromiczne wyciskanie" family). The coach effectively pre-grouped exercises by lesson. There's junk to filter: #ERROR!, ,E, ³sw4ęweee3, 29.07 sen 6h słaba jakość, etc. Let me quickly measure how much the URL signal collapses things and dedupe the doubled URLs before I bring you the approval gate. Created a file, ran a command The URL signal is authoritative and strong — 103 distinct lessons/videos, and the coach himself points whole variant families at one lesson (lesson #5 = 25 phrasings of bench bicep curls; #28 = 19 phrasings of ring/TRX push-ups). I also spotted a filtering issue: some hyperlinked rows are coach notes (Mikołaj Łukaszuk notatki: ...) and sleep logs, not exercises — I'll exclude those. I now have everything needed for the scope approval gate. Here's the proposal: Resolved scope Inputs: 85 .xlsx training-plan files in PLANY TRENINGOWE. Schema (consistent across files): Column A = exercise name; B–E = Tempo/Serie/Powtórzenia/Przerwa; F+ = weekly loads. Multiple date-named sheets per file (training blocks) + empty ArkuszN sheets (skipped). Video URLs live inline in column A text and as cell hyperlinks — overwhelmingly mikolaj.onlinecoursehost.com/.../lessons/N, plus YouTube/Instagram. Proposed dedup key: the base exercise name from column A — with the superset prefix (A /B /C ) stripped, coaching cues after * removed, and URLs pulled out into a separate field. Two grouping signals combined: Shared coach lesson/video URL → authoritative same-exercise grouping (the coach pre-grouped them). Normalized + fuzzy + semantic name matching for the ~82% of entries with no video. Excluded as non-exercises: day headers (Dzień A1), header rows, warm-up placeholders (rozgrzewka), Sen/nie robisz, coach-notes/sleep-log rows, and junk (#ERROR!, ,E, ³sw4ęweee3). Output: consolidated-unique.xlsx with Unique (canonical exercise | video URL | # plans using it | source files) + Merge Audit sheets. Before I run the merge, two decisions materially change the result: Asked Granularity How aggressively should variant phrasings be merged into one exercise? (e.g. 'uginanie łokci na ławce 15 st' / '30 st' / '45 st' / 'hantlami na ławce 30 st') Aggressive (by video) When several names share one coach lesson video, what should the displayed canonical name be? Most frequent variant Decisions noted: aggressive merge anchored on the coach's lesson video, canonical label = most frequent variant. The real payoff here is backfilling videos onto unvideo'd name variants via the lesson grouping. Let me build the full pipeline: improved filtering (drop coach notes/sleep logs/junk), then the four detection passes with URL-anchored grouping, blocked fuzzy matching for tractability. Created a file, ran a command 2,856 → 1,312 groups. The video-anchored families collapsed well (527 URL-names → 85 video groups), but no-video angle/unit variants are still split because fuzzy at 0.82 misses cases like ...na ławce 30 st vs ...45 st (token-set 0.80). Since you chose aggressive, let me add a precise pass that collapses variants differing only in numeric/unit tokens (angles, seconds, weights) — this targets exactly those without over-merging genuinely different movements. Edited a file, ran a command Marginal change — most angle variants also differ in other tokens (added qualifiers, typos). Rather than chase a target number, let me verify quality: inspect the largest merges and detect near-duplicate canonicals that should have merged but didn't, so I can calibrate before the merge-approval gate. Created a file, ran a command This is the critical finding the merge-approval gate exists to catch: raw URL-anchoring over-merges genuinely different exercises, because the coach sometimes demos several movements in one lesson video (or pastes a link onto the wrong row). Look at what got fused: martwy ciąg sumo swallowed martwy ciąg rumuński, martwy ciąg klasyczny, walizkowy martwy ciąg — sumo / Romanian / conventional / suitcase deadlifts are distinct lifts. prostowanie nadgarstków (wrist extension) swallowed uginanie nadgarstków (wrist flexion) — opposite movements — plus a lat-pulldown. wyciskanie półsztangi w półklęku swallowed wiatrak z hantlem (windmill) — unrelated. wznosy kimury + wznosy powella — different raises. Pure name-based passes (typos, word-order, angles) look correct; it's the URL-union that's conflating things. Let me regenerate with video used only to attach the link, not to force a merge, and compare coverage. Edited a file, ran a command Even with URL-merge off, single-linkage fuzzy chaining still over-merges across movement patterns: martwy ciąg sumo chains to Romanian/conventional deadlifts, and worse, wyciskanie poliquina (a press) absorbed wiosłowanie (a row) and rwanie (a snatch). Those are real errors. I'll add a movement-class guard that blocks merges between conflicting patterns (press vs row vs snatch vs squat vs deadlift vs curl vs wrist-flexion vs extension, etc.). Edited pipeline.py +31 -0 Let me check the current file ordering — norm_cache and mclass_cache must be defined before they're used. Read pipeline.py (70–159, 159–203) Now let me wire in mclass_cache and the can_union guard into both the core and fuzzy passes: Edited a file, ran a command The wrist group still fuses flexion+extension — caused by transitive bridging through a class-less variant. Pairwise guarding isn't enough; I need cluster-level class tracking (block a union if the two roots carry conflicting movement classes). Let me refactor to a guarded union with per-root class accumulation. Edited a file, ran a command The wrist flexion/extension are still merged — the guard isn't catching it. Let me debug exactly how those two names are being joined. Ran 2 commands The bridges are typo variants (nadgarsktów — misspelling defeats keyword matching) and a genuine combined entry (prostowanie i ugianie nadgarstków). Perfectly separating wrist flexion/extension is fighting the data — and under aggressive mode, lumping all forearm-cable work into one entry is defensible. The important thing is that the egregious cross-pattern errors are fixed. Let me confirm the press/row/snatch fusion is gone, and generate the full merge-review for your sign-off. Ran Verify press/row/snatch separation Cross-pattern errors are fixed: rowing, snatch, and press are now cleanly separated. The press group stays aggressive (all dumbbell/barbell press angles together) — consistent with your choice. Let me pull the rest of the review (biggest variant families + potential under-merges) to bring you a complete picture for sign-off. Ran Show most-variants groups These aggressive families look right (bicep-curl variants, incline-DB-press variants, bench-press speed/grip variants all collapse correctly). The cross-pattern errors are gone. Remaining imperfections are minor under-merges (e.g. two separate bicep-curl families) and forearm flexion/extension sharing one entry — all visible and reviewable in the Merge Audit sheet. Let me get final distribution stats, then bring you the merge-approval gate. Ran Final distribution stats Merge-approval gate (step 6) After the four passes + movement-class guard, here's the result: Metric Value Unique exercises (final library) 1,457 — multi-variant groups 389 — singletons (appear once/one spelling) 1,068 Exercises with a video URL 227 Training-rows covered by a videoed entry 5,939 / 11,312 (53%) From 2,856 filtered base names (≈2× collapse) What's merged (sample — full list goes in the Merge Audit sheet): ✅ Typos/word-order/diacritics: uginanie/ugianie/uginianie nadgarstków → one ✅ Angle/unit variants: ...na ławce 15/30/45 st → one ✅ Within-movement aggressive families: all incline-DB-press phrasings → one (size 62); bench-press speed/grip variants → one (size 54) 🚫 Blocked cross-pattern fusions: press ≠ row ≠ snatch ≠ squat ≠ deadlift ≠ curl ≠ leg-curl now stay separate Known borderline calls (visible/reversible via the audit sheet): Deadlift variants (sumo/Romanian/conventional) share one entry; squat variants likewise — aggressive