Find EVERY instance of a repeated plan symbol from ONE example — drains, thresholds, fixtures, transition markers: marquee a tight seed_rect around a single instance and the vector linework is searched for every other placement of that same segment cluster. Deterministic geometry, not vision: each placement scores as the length-weighted fraction of the seed's segments reproduced within tolerance_px, under translation plus 0/90/180/270 rotation and mirroring (symbols rotate on plans — both ON by default; turn them off to pin orientation). Score ≥ 0.92 is a match; the 0.75–0.92 band comes back in `withheld` with a reason — a near-match is a question you answer by LOOKING (view_sheet at its `at`), never a silent commit and never a silent drop. RICHER VARIANTS are named, never silent: a placement that reproduces the whole seed but carries >30% extra linework fully inside its footprint (a register against a grille seed — the same outline plus louvers) comes back with its measured `extra` fraction on the row — LOOK at those first, they are the classic mislabel; background lines CROSSING the symbol and coincident duplicate ink never trip this. By default such placements still COUNT, because the contained-seed workflow below depends on supersets matching (seed a bare sub-shape, count the richer symbols that contain it, exclude what you don't mean). Pass `variant_guard: true` when your seed is the WHOLE symbol — grilles, drains, fixtures marqueed complete — and extra-ink placements demote to `withheld` as questions instead of counting; the guard stands down automatically when `exclude` counter-examples are in play, since supplying negatives is manual variant discrimination. The seed's own location is reported in `seed` and never double-committed. Every proposed placement is scored up to a hard work ceiling sized for pathological sheets, and the reply says which it was: complete true means the count is a total; complete false (with candidates.dropped > 0) means the count is a FLOOR — some placements were never scored — so tighten the seed rect around more distinctive geometry rather than trusting it as a total. Marquee discipline: the rect must hug ONE instance — only segments FULLY inside it define the symbol, so a loose rect that swallows wall linework fingerprints the wall, not the symbol. scope "set" sweeps the WHOLE working set, counting on PLAN-role sheets only (the sheet graph decides): a symbol drawn in a detail, legend, or schedule is a reference drawing and never counts itself — which is also how you seed from one: marquee the assembly on the detail sheet and its plan-sheet occurrences are counted while the detail stays excluded (the exclusion disclosed in `skipped`, per-sheet results with per-sheet caps and wall-clock in `sheets`). Scale across sheets: the fingerprint is size-true and is never scale-SEARCHED, so a detail drawn at 1-1/2" = 1'-0" is 12× the size of the same mark on a 1/8" plan — when BOTH sheets have a scale set, the exact ratio is computed from them and the seed is resized before matching (reported per sheet as `scaled`); when a scale is missing, the sweep runs at 1:1 and SAYS so (`scale_assumed`), because an unknown ratio plus a zero count is not evidence of absence. Seeding from a detail/legend/schedule sheet REFUSES outright until both scales are set — that is the case where an unstated ratio silently finds nothing. commit: true (requires condition) commits every match center as an EA count marker through the same path as place_count — the whole sweep (set-wide included) is ONE undo step, each marker carries origin.method "symbol_sweep" with its score, transform, and seed source, and withheld placements are NEVER committed. The SEED instance is not in that count (#296) — in sheet scope it is almost always installed work, so pass commit_seed: true to mint it into the same batch (the reply reminds you whenever a sheet-scope commit leaves it out; ea_total one short of the hand tally is exactly this). The COUNT is scale-free (EA), but matching across sheets of different scales is not — set_scale on the sheets involved is what turns the ratio from an assumption into arithmetic. Counter-examples (#259): drafting reuses one generic shape for different devices — a wall-mounted data outlet drawn as a plain triangle, the flush-floor variant the SAME triangle inside a square, keynote callouts a triangle with a letter in it — so the seed legitimately matches things you do not mean, and seeding more geometry only works where the drawing offers more to capture. `exclude` takes rects around instances you do NOT mean, marqueed exactly like the seed. You never choose a mechanism; the rect's contents decide, because both are the same gesture: a rect holding EXTRA linework beyond the seed rejects placements where that extra linework is present too (the box, the letter), and a rect holding no extra linework of its own is read as the line running THROUGH it — a bare ceiling-grid tile whose grid line a real fixture, drawn over it, would BREAK. That second mechanic is not expressible as a seed: only segments fully INSIDE a rect define a symbol, and background structure is long by nature. Every rejection is disclosed in rejected[] — which negative, what fraction of its evidence was found, and the placement — and NEVER counted in found: an exclusion is a judgement, so look at it and reinstate any you disagree with using place_count at its `at`. A counter-example that holds no instance of the seed, or holds the seed with nothing extra, is REFUSED rather than silently doing nothing. Stroke luminance (#260): a flattened export strips the layer tree and flattens every pen, but the file still STATES stroke color — a black fixture outline over a grey ceiling grid is unambiguous there even when the geometry is identical (two empty 2 ft grid tiles reproduce a 2×4 fixture's outline exactly). luminance_tolerance (0–254) gates on it: a sheet segment only answers for a seed segment when their stroke luminances are within the stated tolerance (Rec. 709, 0 = black, 255 = white; 32–64 separates black from grey without touching anti-aliasing wobble). OPT-IN and disclosed, in the spirit of tolerance_px — omitted, sweeps score exactly as before; stated, the reply's lum_gate says the seed's own luminance band and names every placement the geometry would have committed and the pen did not, so you can LOOK at what a stated gate cost. Prefer geometry (a counter-example, a tighter seed) where the drawing offers it — color is the fallback for exports where nothing else survived. Labels (#308): for a LABELED family — fixtures, tagged equipment, keyed devices — the drawing already names every instance, and the sweep reads those names: a fixture token written beside a placement, or connected to it by a drawn leader line (leader-following arms only on multi-pen sheets, where the annotation pen separates from the work), comes back as `label` + `label_via` on the row, and the seed's own tag rides `seed.label`. Disclosure in both directions, never a recount: a committed match with NO label while the family is labeled was counted on shape alone (measured case: two 0.97 matches that were valve internals, not drains — LOOK at those first), a withheld row carrying the seed's own tag is the drawing vouching for a near-miss (look, then place_count), and a withheld row named a DIFFERENT tag is a sibling fixture answered, not a missed count. After any batch commit, LOOK at what landed — view_sheet {overlay: true} over the swept area — and audit the markers against the drawing before trusting the EA total. Coordinates are image px at render scale 2.0: PDF pt × 2, origin top-left, y down (the browser canvas's native space). Sheet payloads carry dims in both px and pt.