[guijacquemet]

Adds docs/SPEC_WRONG_SCIENCE.md (docs only, no code) and aligns related docs with the maintainers' decisions. Known-answer fixtures are data, meant for conformance.json key wrong_science and the conformance suite (#87). Spec status: v1 is 2D; N-D and the QuPath resolution choice follow in Wave 2.

Themes: 1 coordinate frame, 2 calibration, 3 retention, 4 run identity, 5 async ChoicesFrom lookups, 6 number limits.

What changed in the latest update

• Retention: hosts never prune and never delete under results/ or runs/. Hosts only write .lc-owner.json and .lc-done.json, then start labconstrictor-tools prune in a detached subprocess. Only Tools prunes (labconstrictor_tools.retention, CLI prune). GUI-only users get cleanup through that CLI. Removed: host-side prune, Java line-by-line port, cross-language lock text. RET-01..07 are now "prune eligibility" fixtures against the Tools reference; RET-08 is CLI-only concurrent prunes; new HRET-01..04 assert hosts never delete. The FileOut move before temp-folder deletion is kept (not pruning). Windows note: junction via mklink /J, plain symlinks need admin.
• Calibration: _calibration is per axis (x, y, z, t each {value, unit, source} plus overall provenance); phase 1 fills x, y. New fixtures: TIFF INCH 72/96 dpi = unknown, micro-sign spellings (U+00B5, U+03BC, U+FFFD), Angstrom. Anisotropic rule (>1e-3) stated as a behaviour change with a release-note line.
• Coordinate frame: frame is {source, shape, axes_offset (t,z,y,x), plane, downsample, transform}. One placement formula: c = x0 + (x + 0.5) * downsample. Results with no t/z land on the exported plane (not plane 0). QuPath whole-image and viewport export use the same plane. Fixture CF-14 (downsample 2 and 4) added; Wave 2 fixtures needed are listed in the spec.
• Number limits (new section 6): form integers up to +-(2^53-1), floats +-1e15; beyond is refused, never clamped. Fixtures NUM-01..05 per host.

Related docs edited for consistency

• SPEC_WAVE2.md: 1.4 calibration rules rewritten (unknown, blank field, run blocked, no default stand-in, per-axis _calibration; example drops the = 1.0 defaults); 1.5 and open question 6 now say "exported plane", not plane 0; calibration_prefill case updated.
• HOST_PARITY.md: C19 (bare scale = unknown, not assumed um), A06, A07, C01, D05 (limits), L02, L03 (hosts never prune), fix rows FJ-4, QP-18, NB-7, NB-1, QP-2c, and the conformance.json sketch (form_integer_max, unbounded_float).
• HOST_FEATURES.md: PixelSizeOf row. OPERATIONS.md, AUTHORING.md: cleanup is done by labconstrictor-tools prune.
• Consistency grep over the specs, HOST_*, PROTOCOL, OPERATIONS, AUTHORING for assumed / prune / flock / RET-08 / plane 0: remaining hits are intentional.

Section 8 of the spec lists decided items and the working decisions marked "default, not yet confirmed". Not merged; for review.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Y2SN8SuEakTdiWy3F2WUqN