# BidBrain project status, part 18: The purchases module rebuilt as a working list (27 August 2026) Moved unchanged from CLAUDE.md on 2026-09-18. The standing rules stay in CLAUDE.md, which lists every part of this log. - THE PURCHASES MODULE REBUILT AS A WORKING LIST, 2026-08-27 (Steven's own build, the big one from the Gemini meeting task list: "rebuild the purchases module UI", designed across several mockup rounds he signed off one by one, then "lets build it"). The purchases page is no longer a passive report of what the platforms said, it is the live working list for every car between winning the bid and landing on site, one table, Airtable style inline editing, a resizable record drawer, a real per car timeline, chips, cancellations and a journey with a Check in that archives the car. Built on Fable 5 per the model recommendation for this scale of change. NOT yet committed or pushed, per Steven's own release discipline, and MARK MUST REPEAT THE DATA GROUNDWORK ON HIS OWN MAC when this lands (see the memory file purchases-rebuild-mark-handover.md and the Mark notes below). - THE DATA GROUNDWORK, applied to Steven's live database first (backups taken: bidbrain.db.backup-20260827-163058, -prebid, -phaseA): 88 platform cancelled sales flagged per purchase (not reg keyed, see the Puma below), 104 old finished cars moved to stock_status "In stock" (the archive), leaving his real 21 car working list; the BMW X1 corrupted row fixed (SD66XWV, real reg and name recovered); the Ford Puma double sale resolved (listed and won on BOTH platforms the same day, the Motorway one cancelled, the Carwow one live, which is exactly why cancellation is now recorded on the PURCHASE ROW, cancelled_at/reason/source, never derived from the reg keyed walked_away table, whose reg only identity cannot say WHICH sale died); 165 missing Motorway winning bids recovered by inverting the fee band arithmetic (all_in = bid + fee(bid) x 1.2, the recovery self verified against all 30 rows that carried both figures before any write, 30 of 30 exact, SA18BSO honestly left empty as unrecoverable, recovered rows written payment_estimated=1 so a later real breakdown read still replaces them); GBP 64,542 that had been sitting inside "price" reclassified as fees; the CompMatch comp inflation this caused corrected (about GBP 390 per car on 116D, A Class, Qashqai comps). - CANONICAL MONEY, ONE FUNNEL: db.list_purchases now derives every money figure once, per row: bid_original (winning_bid, else the platform price, else None for a Motorway row with no breakdown, NEVER the all in figure standing in for a bid), chip_total, car_price (bid less chips), effective_fee/effective_vat (the last chip's frozen re band, else as read), fees_total, all_in (car_price plus fees, falling back to the platform's own total for an old row). The page and the DealerKit push only ever read these, never re derive from raw columns. - CHIPS ARE FROZEN EVENTS, NEVER LIVE ARITHMETIC: purchase_chips stores amount, reason, bid_before, and the new fee and VAT computed ONCE at chip time (serve._record_chip, against the db.fee_for band tables) and frozen, never recomputed against a table that may have changed. Motorway re bands (Right Drive is on Motorway Pay and Instant Billing), Carwow never does, fees follow the ORIGINAL price there (Steven's final confirmation, matches FE68CKV's real invoice), both per dealership settings (purchase_flow.fee_rebands). Chips lock the payment field block in record_purchases so a later platform sync can never overwrite a chipped price back to the original; a platform read that MATCHES the chipped figure logs a one time timeline notice instead. Multiple chips per car stack; the platform's locked price is the ONE piece of locked platform data a chip is allowed to move, automatically, shown in the timeline as original to new (Steven's explicit design). - THE TIMELINE: purchase_events, one row per change, through a single mutation funnel (db.set_purchase_field with its editable whitelist, record_chip, cancel_purchase, check_in_purchase, add_purchase_comment), every entry stamped with who and when, comments included (Steven: "every change to any vehicle record should be shown in the timeline and timestamped including all comments and status changes"). The page has no private write path, every edit goes through /api/set-purchase-field, which also refuses a dropdown value that is not one of the configured options. - THE PAGE (render_purchases, rebuilt): four views (Due in, the working list, everything not In stock and not cancelled, default sorted soonest collection first; In stock, the archive, checking in IS archiving; Cancelled, every fallen through sale kept viewable with its reason; All), independent of the platform tabs. The old June 2026 date pre fill is gone, the views make it redundant. Columns per Steven's Airtable: Status, Video, Person collecting, Paid for, On finance, Notes (renamed from Due in status, his ask), Collection arranged, Location, Car price, Chipped, plus the existing Retail est., Service history, Logbook (V5 scans plus a logbook_state pill in one cell), DealerKit. Every column sortable (SORT_KEYS extended, each key returns exactly what its column displays), every column hideable per Table and per Record separately via the Fields menu, saved server side (purchase_columns) so it holds on every device. Inline editors: a floating option menu for dropdown columns, a floating input for text, number and datetime, both committing through saveField. Reg and Vehicle open the record drawer, no row numbers, no vertical separators (his ask). An Add car by hand modal for a car bought outside the platforms (a local trader), platform "Manual", nothing ever overwrites it, marked with an M tag. - THE RECORD DRAWER, the approved journey header design (his explicit mid build instruction that it must be in the build): the stepper above the tabs (Bought, Video check when purchase_flow.video_stage is on, Steven yes Mark off, Collection which Paid for closes, Arrived which Check in closes), a Next up card saying the one thing to do next, collapsible stage cards with their own editable rows, tabs Car details / Timeline / DealerKit, an always visible comments box, a drag handle whose width persists per device (localStorage bb_drawer_w), and a three dot actions menu holding Record a chip (any point, more than once) and Cancel this sale (any point, reason type from the configured list plus required free text, kept for the proof pack). The video stage closes on the LAST option of the dealership's own video_status list, so the order in settings is meaningful, not just cosmetic. The DealerKit tab is a read only summary of the last Check DK record including a price mismatch warning when DealerKit's stored purchase price disagrees with the canonical car_price (a chip moves BidBrain, never DealerKit, on its own). - MILEAGE HARD BLOCK, NO ESCAPE HATCH ANYWHERE (Steven: "NEVER ALLOW Dealerkit to manually guess the mileage"): the old allow_missing_mileage confirm-through flow is REMOVED from dealerkit_write, serve.py, purchases_run.py, daily_run's bulk pass and the page (the Upload button is disabled outright with "Enter the mileage first"); a car with no mileage can never create a DealerKit record from any path. - SETTINGS: a new Purchases tab on the settings page edits every dropdown column's own answers (add, reorder with real up/down, recolour, flag default/finished/retired, remove) plus the journey flags (video stage on/off, fee re bands per platform). Saves through its own /api/purchase-page-settings (the full /api/settings route needs the whole buying rules payload this page never holds). A REAL IN USE GUARD server side: removing an answer still carried by real cars is refused naming the count ("'Due in' is still on 109 car(s)"), retire it instead, a retired answer stays readable on old cars and is never offered again (openSelMenu filters them, videoDone ignores them). Fee band tables themselves stay code only, deliberately, they only change when a platform publishes new ones. Steven's own seeded config: person_collecting Relay Network/Carl/Collected in house, video order Awaiting confirmation, More info needed, Video requested, Video received, Video signed off (signed off is the gate), location/aucstatus/duein columns off, video_stage on, rebands Motorway yes Carwow no. - BUGS CAUGHT BUILDING IT: the settings panel's save button was wired with a parse time addEventListener but lives inside the cloned form template, so it was null at parse time and the whole settings page stuck on Loading (the exact template-clone timing class of bug this page has had before), fixed with an inline onclick; a mangled ternary in the chip cell caught before it ever rendered; the settings page needed its own copy of the pill palette (it shares no stylesheet with the purchases page), hex for hex identical. - VERIFIED LIVE END TO END in a real browser against Steven's real data, not just unit tests: the 21 car working list, drawer open on the real BMW X1, video status picked from his real options (Next up updated, table pill updated, timeline logged the change), a comment posted (landed in Comments AND the timeline), a chip recorded (live preview showed 7,500 to 7,250 with the Carwow fees-follow-original hint, header re read "after GBP 250 chipped"), a collection date committed through the floating editor (the car re sorted to the top of the working list), In stock view showed exactly the 104 archived cars and Cancelled exactly the 88, the Fields menu round tripped a column off and back on (confirmed in the database both ways), the Stats dropdown read the canonical figures (21 showing, GBP 141,649.40 all in, chip and cancellation rollups), and the in use guard refused a real removal. EVERY test artifact from that browser pass (the chip, the comment, the video status, the collection date, their timeline events) was then removed from the real database and verified gone, Steven's data is exactly as it was. 600 tests pass (the ~30 new Phase A checks included), node --check clean on both generated pages, no leftover template tokens. - DRAWER SPACING, "EVERYTHING ELSE" AND WIDE WIDTH RESPONSIVENESS REDESIGNED, same day, straight after (Steven: "there are some clear spacing issues on the car overlay... the everything else section needs re thought out, it looks like its part of the car journey at the moment, and its not... when you make it bigger it creates too much space in width between the title and the data"). Looked at the real rendered drawer at 940px before changing anything, which surfaced a genuine BUG beyond the spacing: the header's three dot actions button was absolutely positioned inside an inline span wrapper, so its right offset resolved against the zero size span rather than the header and the button rendered stranded under the platform pill. Header rebuilt as a real flex layout (identity row: reg plate, platform pill, manual tag, actions pushed right; the vehicle name on its own full width line beneath), no absolute positioning at all. The other three fixes, all live verified at 520px and 940px: (1) Field rows (.sgr) changed from space-between (which flung the value to the far edge on a wide drawer, the exact complaint) to a fixed label column grid, the value always sits right next to its label. (2) A CSS container query on the DRAWER's own width (container-type inline-size, not a viewport media query, since the drawer is resized by dragging): past 640px the stage body and section rows lay out two up so the extra width carries data instead of stretching one thin column, with the money block, notes, the closing row and the Check in button spanning both. The stepper is capped at 560px and centred so a wide drawer keeps it composed. (3) The Bought stage's money block was a flex strip of up to 8 equal columns (squashed narrow, sparse wide), now a proper receipt, one figure per line in a quiet bounded box, Car price and All in ruled off and larger. "Everything else" is no longer a .jstage at all: a new flat .dsec section (small caps heading "Other details", no dot, no chevron, no card border, unindented) so it reads as reference data under the journey, never a fifth stage; the drawer's DealerKit pane got the same flat treatment for the same reason. Gutters standardised on 20px throughout the drawer (header, stages, next up, cancel banner, timeline, comments). 600 tests pass, node --check clean, verified in the real browser at both widths against the real BMW X1 record. - FINE TOOTH COMB AUDIT AND POLISH PASS, same day (Steven: "there are things broken, and there are things missing from the build that we discussed... the usability is quite shocking... do a user experience test and find faults that are obvious and not so obvious"). Run as a real hands on UX test, not a code review: two throwaway manual cars (TEST111, TEST222) driven through the WHOLE real lifecycle in the browser, add by hand, drag resize, every editor type, video to signed off, person collecting, paid for, a chip, check in, and a full cancellation, every screenshot inspected, then both cars and every trace of them (events, chips, walked_away, the test saved view) deleted from the real database and verified gone. What the test PROVED working: the drag resize, the whole journey progression, the chip preview and recording, the cancel dialog, search, sorting, the Fields menu round trip, check in, comments landing in both places. BROKEN, found and fixed: (1) A manual car's platform pill was INVISIBLE, white text on nothing, .plat.manual had no CSS anywhere (both the purchases and lost bids pages share the rule, both fixed). (2) A cancelled car's record lost its whole bought history, the Car details pane showed nothing but Other details, the money and dates gone; it now keeps a flat Bought section (platform, date, mileage, the full money receipt) above Other details. (3) Every cancelled car's Status read "Due in", nonsense; the Status cell and the record's own Status row now say Cancelled in red on a cancelled car, and are not editable there. (4) The video gate followed "the last option in the list" rather than the option actually FLAGGED gate in settings, so reordering the list in the new editor would have silently moved the gate; videoGateValue() now reads the flag, last option only as a fallback, and the stage's own note names the real gate option. (5) The default flags on dropdown options did nothing at all: an untouched Video cell showed "–" when the car genuinely IS "Awaiting confirmation"; an empty cell now shows the default flagged option's pill faded, titled "The default, not set by hand yet". (6) Escape closed nothing; it now backs out one layer at a time, editor or menu, then dialog, then the drawer, leaving the photo gallery's own handler alone. (7) Timeline entries showed raw stored values ("1" for paid, a bare ISO datetime); paid_for reads Yes/No, dates and mileage read their human shape, display only. (8) Moving a car to In stock straight from the table's Status cell left the record's Arrived stage offering a Check in for a car already archived; it now says plainly "In stock, handled in DealerKit from here" with no fake arrival stamp. (9) Table dates were raw ISO (2026-08-26); Bought and Due in date now read 26/8/2026 like everywhere else on the record. MISSING FROM THE DISCUSSED BUILD, now built: (10) Column rollups at the foot of the table (Steven's explicit "rollups at the bottom of the columns"): a real tfoot with the car count, Car price total, Chipped total, Retail est. total and a paid count, built from the SAME column visibility calls as the header and rows so they can never disagree. (11) Saved views (Steven's original "...and saved as a view"): a Views menu next to Fields saves the page exactly as it stands (view tab, platform tab, search, dates, sort) under a name, apply with one click, delete with a confirm; stored server side per dealership (settings key purchase_views, its own validator in serve.py, only known criteria keys survive), so views hold on every device. 4 new tests. (12) Per view column tailoring, the not so obvious usability fault: the Cancelled view used to show every working list column (Video, Person collecting, Paid...) and pushed the cancellation reason, the one column that view EXISTS for, off the edge of the screen; In stock had the same problem milder. VIEW_HIDE now trims each view to the columns that mean something there, the Cancelled view reads Photo, Reg, Vehicle, Platform, Bought, Notes, Car price, Chipped, Retail est., Cancelled, all on screen with no sideways scroll, and the Fields menu's own choices still apply within that. Verified live end to end after every fix, 604 tests pass, node --check clean, no leftover template tokens, and the real page confirmed back to Steven's own 21 car working list with zero test data left behind. - THE DRAWER REBUILT AS RECORD LEFT, JOURNEY RIGHT, same day (Steven: "car details need to be its own section and the actual flow needs to be its own section too... you have fields in places that doesnt even make any sense", then, from three mockups, "i actually like c, but lets force it to always be wide so its always side by side. then always show car details on the left and move the tabs to the right so the 3 tabs are journey - timeline - dealerkit", then "in the dealerkit section it needs to have all the dealerkit sync options showing whats missing and whats not synced", then "that looks perfect"). Designed as three interactive mockups on an artifact first (Record and Journey Concepts), built only after his sign off. The underlying rule that fixed the confusion: EVERY FIELD HAS EXACTLY ONE HOME. The record owns the facts, the journey owns only what moves the car forward, and a journey row may quote a fact but never edits it. The built layout: the drawer always opens wide enough for the split (940px default, drag clamped to at least 780, a remembered narrower width from the old drawer is clamped up on open), header (identity, stepper, money line) spanning both columns, then a permanent CAR DETAILS column on the left (four groups: Car, Money receipt, Collection and logistics, Notes and documents, pencil marks on every editable row) and a right panel carrying the three tabs Journey, Timeline, DealerKit. The old expanding stage cards, the "Other details" section and the Car details tab are gone entirely (markup, JS and CSS all removed, not disabled). The container query now handles only the genuinely narrow case, a phone, by stacking the right panel under the record. The Journey tab: the Next up card plus a vertical rail, one row per stage with ONLY its single closing control, the video status pill, the Paid for tick, the Check in button (disabled with "Finish the earlier stages first" until it is the current stage, a real [disabled] style added after the first live check showed it rendering full strength). Collection's row quotes the booking and collector from the record, read only. The DealerKit tab is the FULL sync panel, not a read only summary: every item (Car, Purchase price, Auction fee, Delivery, Indemnity, Photos, Documents, Service history, V5, Retail price, Due in stage, Due in date) in one of four honest states, a green tick with the real figure when on DealerKit, a red ring when missing there, a blue ticked box when BidBrain has it to push (the gap spelled out, "£7,500 on DK, £7,250 here"), dim "Not charged on this purchase" when not applicable, grey rings before any check has run. A price mismatch after a chip arrives pre ticked with the warning underneath. Upload ("Upload N ticked items to DK", the count live from the boxes) and Check DK record both run from the panel through the same endpoints and progress polling the table row uses, and loadPurchases now repaints an open drawer, so a push or check finishing shows in the drawer too, not just the table behind it. The mileage hard block holds here identically, the button reads "Enter the mileage first" and disables outright. Verified live end to end on the real BMW X1: the split renders, the video pill opens its options from the rail, the DK panel showed the car's true never checked state with the correct "Upload 3 ticked items" count, the disabled Check in reads disabled, and Steven's own data was left untouched throughout. 604 tests pass, node --check clean, no orphan handlers, no leftover template tokens. - COLUMN DRAG REORDER, COLUMN RESIZE, AND THE DROPDOWNS MADE HONEST, same day (Steven: "you still cant drag and drop the order of the columns? and you still cant resize a column? this was supposed to be in the build? but you missed it... also when you click a dropdown you cant click off it properly, it also isnt clear to the user that they are actual dropdowns"). He was right on all three, reorder and resize were part of the Airtable brief and had been sitting on the deferred list. (1) The whole table (rows, header, footer, and a new colgroup) was refactored from a hard coded column concatenation to ONE ordered list of column keys, so all four can never disagree. Each column's cfg (the same purchase_columns settings the Fields menu writes) gained optional pos (drag order) and w (a dragged width), validator widened in serve.py, saved per dealership like everything else on this page. Photo, Reg and Vehicle stay fixed first, the Cancelled column pinned last, everything between drags: real HTML5 drag and drop on the headers with a drop side indicator, reordering over the FULL movable list so hidden columns keep their relative places, the Fields menu now lists columns in the same configured order. (2) Resize: a grip on every header edge (Photo excepted), pointer drag sets the col element's width live and saves it on release (48 to 700px); a never resized column keeps the browser's own natural width so nothing jumped when this shipped. Honest limit: the table keeps auto layout (fixed layout plus overflow clipping would have clipped the DK rollover popovers), so dragging a column NARROWER than its own content stops at the content's minimum, widening always works. A grip press can never fire the sort, and a drag started on the grip never starts a column drag. (3) Dropdown dismissal was genuinely broken in three ways: clicking the same cell again blinked the menu shut and straight back open (now a real toggle, suppressReopen), several cells stop the click to protect their own action so the plain document click closer never ran (now a capture phase pointerdown closes menus on ANY press outside them, which stopPropagation on click cannot swallow), and scrolling left the position fixed menu hanging in space (any scroll now closes menus and commits an open text editor through its own blur). Opening Fields no longer leaves Views open either, closeMenus finally knows about all three panels. The state lives ABOVE the functions that use it, the third time this file nearly hit its own documented temporal dead zone trap. (4) Affordance: every dropdown cell (table, drawer record rows, the journey rail's video pill) now carries a small chevron caret, darkening on hover; the table's text and date cells show a pencil on hover, matching the drawer's own. And a sorted column can finally be UNSORTED, a third header click clears back to the view's own default order, there was previously no way back at all. Verified live: the same cell toggle proven with real clicks, reorder and resize proven by dispatching real DragEvents and PointerEvents through the actual handlers (this pane's synthetic mouse drags cannot trigger native HTML5 drag, the same tool limitation as the datetime segments), Car price genuinely moved before Bought and persisted to the settings table, the Vehicle column resized 250 to 310px and saved; every test artifact (positions, widths) was then stripped from the real settings with Steven's own column visibility choices verified intact. 604 tests pass. - RESIZE ACTUALLY FIXED, AND THE TABLE'S TITLES PINNED, same day (Steven, with a screenshot: "everything in this screenshot including the titles should be fixed and not affected by the scroll of the table. also the resizing the columns still doesnt work"). The resize failure had a real root cause the first verification MISSED: the native drag's SOURCE is the draggable th itself, so dragstart fires with the th as its target, never the grip, the grip check in dragstart could not match, and a real mouse press on the grip became a column drag that cancels the pointer stream dead. The synthetic pointer test had passed because it never exercised the native drag machinery at all, a false pass worth remembering: a synthetic event test proves the handler, never the browser's own competing gesture. Fixed belt and braces: the grip's pointerdown suspends the th's draggable for the length of the gesture (restored on pointerup AND pointercancel), captures the pointer so move and release always arrive, and dragstart additionally refuses while a resize is in flight. The titles: the table now scrolls INSIDE its own box (the wrapper capped to the viewport by sizeTableArea, measured, not a guessed calc, since the run bar and notices above come and go), so the view buttons and toolbar never move, the column titles are position sticky at the box's top and the rollup footer at its bottom, always visible. border-collapse had to become separate, collapsed borders shed from sticky headers in Chrome as they float. A REAL MISTAKE MADE AND PUT RIGHT during this: the test cleanup, stripping the widths and positions the verification pass itself had saved, assumed EVERYTHING it found was its own artifact, but Steven was using the feature live at that exact moment and most of what it found was HIS OWN first real column arrangement (Status first, then Bought, then Platform, his own widths), wiped along with the test values. The exact values had been printed before the wipe, so his whole layout was restored verbatim and verified equal; the one ambiguous value is the Video column width (210px, which may include 40px from a test drag on that same grip), one drag for him to correct if so. The lesson, recorded so it sticks: on a LIVE system the user may be acting between verification steps, so a cleanup must only ever remove values it can positively attribute to the test (captured before and after), never "whatever is there now". 604 tests pass. - RESIZE FIXED FOR REAL AT THE THIRD ATTEMPT, same day (Steven: "the resizing still isnt working, fix it properly!"). TWO more real causes found, each masked by the last: (1) The grip was nearly unhittable: 9px wide with 4px hanging into the NEXT header, and since every sticky header is its own stacking context the neighbour painted over the overhang, leaving about 5 invisible pixels. Reworked: a wide right grip fully inside its own header PLUS a left grip on every header that resizes the PREVIOUS column, so the whole strip around a border grabs from either side, with a divider that shows on header hover so it can be found at all. (2) The deepest one: even when the handler ran and saved, the table used auto layout with width 100%, so the browser treated the col widths as SUGGESTIONS and redistributed them, the saved number changed nothing on screen. Every earlier "pass" had checked the style attribute or the config, never the RENDERED width; Steven's own drags were genuinely saving figures (a 355px Vehicle width sat in his config from an attempt that visibly did nothing). Fixed with table-layout fixed, a width for every column (the dragged figure or a sane default), the used table width being the larger of 100% and the column sum so it still fills when narrow and scrolls sideways when wide, and overflow hidden on the text cells EXCEPT the popover hosts (Auc status, the money columns, DealerKit) whose rollovers would be clipped with them. Proven the right way at last: a real CDP mouse drag on the grip (the same input pipeline as a real mouse, not a synthetic PointerEvent) moved the RENDERED width 310 to 370 and saved it; config snapshotted before the test and restored verbatim after, the lesson from the wipe earlier the same day applied. The standing verification lesson now twice earned: check the outcome the user sees (rendered layout), never the mechanism's own bookkeeping. 604 tests pass. - THE CELL EDITORS MADE INLINE, THE DATE PICKER OPENS ITSELF, same day (Steven: "the collection arranged date picker doesnt work as expected, it should just be the date and time picker, no typing, and it should be inline, its behaving funny. also the same can be said for notes"). The old editor was a floating box appended to the body and positioned NEAR the clicked cell: the datetime one landed as a segmented typing field whose tiny calendar icon needed a second, precise click, the notes one was a floating textarea where Enter made a newline instead of saving, and both drifted or dismissed oddly around the new scroll and pointer handling. Rewritten to edit IN PLACE: the input replaces the clicked cell's own content (a record row edits in its value slot so the label stays put), a date field calls the browser's own showPicker() immediately so the real calendar opens on the first click with no typing, notes save on Enter (Shift Enter for a new line), Escape backs out without saving and without falling through to close the drawer, and clicking anywhere else saves through blur. closeFloaters now BLURS an open inline editor rather than removing it, so no dismissal path can ever throw an edit away, and the fixed table layout means the inline input fits its cell exactly. Verified live on a throwaway manual car (TEST333, deleted with its events afterwards, zero traces): the datetime editor rendered inside its cell focused with the native picker up (the popup held focus against synthetic keys, itself the proof it opened), a committed value saved and rendered as 29/8/2026, 2:30pm, real typed notes committed on Enter, and Escape discarded a poisoned value without touching the saved one. One tool note for future sessions: this pane's key action does not reliably deliver Enter or Escape to the page (typing works), so keyboard paths are verified with dispatched KeyboardEvents instead. 604 tests pass. - LONG NOTES NO LONGER KICK YOU OUT, PENCE ON EVERY PRICE, UK DATES, same day (Steven: "notes is broke, you can only type so many characters and then it kicks you out and wont let you edit them... prices across the app should have two decimal points for pence... dates are in american format, this is a uk app"). (1) The notes kick out had a precise mechanism: the capture phase scroll listener (added earlier the same day to dismiss floating menus on scroll) caught the textarea's OWN internal scroll, which fires the moment the text wraps past the visible box, and blurred the editor mid word; opening an already long note scrolled to the caret instantly, so editing was killed on arrival, exactly "wont let you edit them". The listener now ignores scrolls originating inside an inline editor, the notes box starts two rows tall and GROWS with the text (capped at 220px, then scrolling inside itself, safely now), and the cell being edited lifts its own overflow clipping for the duration so the grown editor is never cut off. Proven live with a 217 character note typed in one go, editor open and focused throughout, saved intact. (2) Money: pounds() on both the purchases and Didn't win pages now always shows pence (en-GB, exactly two decimals), and every raw toLocaleString money spot (the chip cell, the chip receipt lines, the footer chip total, the header's "after £X chipped") went through a shared poundsNeg/pounds instead, so £4,820.8 style one decimal figures are gone. The cockpit's own valuation figures (max bid, retail) deliberately stay whole pounds, they are estimates that never carry pence; say the word if those should change too. (3) Dates: fmtDT and fmtStamp now zero pad (27/08/2026), the drawer's money line no longer shows a raw ISO bought date, and the Didn't win page's own date column (raw ISO, year first, the likeliest "american format" sighting) now reads DD/MM/YYYY via its own ukDate helper, which a placement check caught being defined on the WRONG page's script first (each page carries its own script, a helper on one is undefined on the other), fixed before it shipped. One honest caveat told to Steven: the date picker's own segment order inside the native input follows Chrome's language setting, not the app; every date the APP renders is UK. 604 tests pass, both pages' generated JS node checked, the throwaway test car deleted with zero traces. - PENCE ONLY WHEN REAL, AND THE NOTES CELL BACK IN LINE, same day (Steven: "pence should never show if it is .00... the notes are working better but they are now out of line with the rest of the table"). pounds() on both money pages now shows pence only when they are non zero: £7,500 stays clean, £4,820.80 keeps both digits, proven against the page's own live function (£7,500 / £4,820.80 / £1,234.56 / £0.50 / –). The misalignment had a precise cause: the .notescell CSS dated from the OLD page where it styled an inline SPAN, and its display:inline-block, applied since the rebuild to the td itself, stopped that cell being a table cell at all and knocked the whole row out of column. Trimmed to ellipsis and font size only; verified live that the notes td computes as table-cell again and every one of the 18 cells in a real row sits within 2px of its own header's left edge, zero misaligned. 604 tests pass. - A TICK TO FINISH A NOTE, AN HONEST EMPTY SERVICE HISTORY, AND DOCUMENT PHOTOS ON THE RECORD, same day (Steven: "when you start writing a note there needs to be a tick button at the end... currently its awkward to click out of it. if a car has no service history, instead of being blank it needs to say No Service History. also in the car overlay you should be able to see the uploaded service history pics as well as the v5 pics the same way you can in the table view"). (1) The notes editor grows a green tick button under its right corner; its pointerdown is prevented so pressing it never blurs the textarea first, the click is the one clean commit path (Enter and clicking away still save, Escape still discards). (2) _docCell gained an empty state parameter: the Service history cell reads a small red "No service history" instead of a dash, in the table AND on the record, never a blank. (3) The record's Notes and documents group shows the REAL uploaded service history and V5 photos as clickable thumbnails through the exact same _docCell and shared gallery the table row uses (mediaUrl for Motorway's urls, localDocUrl for Carwow's local files), replacing the old bare page counts. Verified live end to end: a throwaway car proved the tick commits ("Saved with the tick button" persisted, editor closed), the empty state shows in both places, and a real documented car's record row opened the shared gallery from its own thumbnail; the test car deleted with zero traces. 604 tests pass. - STILL TO DO, deliberately deferred: a fee band editor UI, CompMatch deliberately keeps using the pre chip bid as its comp (a chip reflects condition found, not market level). STANDING REMINDERS: Steven's own NEXT JOB after this is the DealerKit half of Check in (pushing the arrival into DealerKit); dealerkit_write.set_purchase_price has never been watched live, its first real use must be watched not trusted; the platform cancellation proof pack is a future feature. FOR MARK when this lands: he must repeat the data tidy up on his own database (his own cancelled/archived split, his own bid recovery IF his account's arithmetic checks out, the Motorway Move caveat in the memory file means a blind rerun of Steven's recovery could be wrong AND pass its own check), and he should be asked his own dropdown answers (all columns), whether he wants Location and the Due in date column, what to do with Auc Status, and told video_stage defaults off for him; A4C, Dealer Auction and DealerWay purchases remain his own later build.