Improve curve-fit reports, uncertainty handling, and results layout

[paulnord]

Curve-fit results now provide unit-bearing spreadsheet exports and experimental-physics fit statistics while keeping equal-weight least squares as the default. Expanding the report or changing the optional common measurement uncertainty does not change the best-fit coefficients.

Copy Fit Report now defaults to a compact classroom summary: model, equation, variable units, coefficients and errors, and the statistics shown on screen. The adjacent ... menu offers Copy Full Fit Report for the detailed analysis. Both retain full-precision spreadsheet cells.

The full report explicitly identifies X/Y variables and units, observations, editable parameter count, independent parameter rank, rank-aware degrees of freedom, residual statistics, centered R Square, chi-square, and one full-precision parameter table. For example, quadratic position-versus-time coefficients have m/s^2, m/s, and m units, following KnownPolynomial's verified descending-degree parameter order. UserFunction parameter units are not guessed.

DataTable exports obtain units from existing metadata, not column-name parsing. Raw and formatted copies share unit-bearing headers while retaining their numeric serialization and delimiters. ASCII expressions replace display superscripts, and copy/paste does not append duplicate unit suffixes. Unitless columns retain their previous subscript normalization.

The compact optional uncertainty controls support:

• Estimated from residuals (default): sigma_y = sqrt(SSE/(n-rank)), chi-square = df and reduced chi-square = 1 by construction when defined; Q is N/A. The report explicitly states this is not an independent goodness-of-fit test.
• A specified positive constant sigma_y: chi-square = SSE/sigma_y^2, reduced chi-square = chi-square/df, Q = the chi-square survival probability. The supplied scale is never recalibrated. Invalid input remains unavailable rather than falling back to estimated noise.
• Fractional/custom pixels when a host supplies a valid direct-position conversion. OSP exposes a session-only FitMetadataProvider API; companion Tracker draft PR #74 supplies its existing calibration metadata. No project format change or general dimensional-analysis engine is introduced.

Coefficient standard errors retain Tracker's Bevington Eq. 8.13 profile/refit curvature method. Supplied sigma changes its scale; it does not replace the method with Jacobian covariance. Redundant parameters retain N/A errors and an explicit identifiability warning. Perfect residual-estimated fits report undefined chi-square rather than 0/0; a supplied positive sigma permits zero chi-square and Q=1.

Centered R Square remains descriptive variance explained, with the existing caution about Excel's uncentered convention for some through-zero fits. Rank is based on normalized derivative columns with tolerance 1e-7; nonlinear interpretation is local. Valid classical polynomial ANOVA remains supplemental. No coefficient p-values or confidence intervals were added.

Earlier improvements in this PR remain: two significant uncertainty digits with matching coefficient precision, full-precision values for computation/export, immediate fixed/free parameter updates in Autofit, manual-mode preservation, a Copy Fit Report button, context-menu gestures that preserve Autofit, and compact fit information/statistics with stable layout during selection and resizing.

Validation of this revision:

• macOS desktop: 592 assertions across precision (19), report (118), popup (17), Data Tool layout (234), constraints (40), and physics/units (164); all pass. The fixture cleanup now clears its disposable tab’s changed flag before disposal to avoid a modal save prompt that previously caused intermittent timeouts.
• Chrome, Firefox, and Playwright WebKit: 341 computational assertions per browser (19 precision, 118 report, 40 constraint, 164 physics).
• Browser UI: resizing and point selection preserve layout; fixed parameters and supplied sigma update reporting correctly; coefficients stay unchanged when sigma changes; actual compact and full report copy gestures work. Chrome clipboard readback matches the expected TSV exactly. Firefox/WebKit clipboard contents were not read back.
• The companion Tracker adapter builds against the new OSP classes and passes 17 calibration/column-identity checks; its four source files transpile with SwingJS. OSP changed classes and regression fixtures also transpile successfully.
• Earlier revisions passed native Windows/Linux/macOS CI, but Windows/Linux were not rerun for this revision. These are targeted curve-fit tests, not comprehensive Tracker/installer/video-codec validation or native Safari testing.

Statistical definitions, metadata integration, limitations, and test instructions are in test/fit-report-physics.md and test/README.md. Dataset drawing error bars were inspected; a future per-point uncertainty column needs explicit selection-aligned mapping and is outside this change.

Uncertainty controls now explicitly label the default “Unweighted - estimate from residuals,” provide a wider separate value row, and distinguish pixel presets from physical units. Physical-unit starting values use one calibrated pixel (or residual scatter without calibration); switching supplied modes converts the current scale. Desktop layout tests check selector and value-field visibility and fractional-value entry width.

The display uses a single row for R-squared, SSE, and Residual SE; observation counts and degrees of freedom remain in copied reports. A scrolling fit panel preserves drawable graph space in small windows. The optional host position-versus-time metadata enables copied line-fit velocity and quadratic-fit velocity at t=0 and acceleration, with errors propagated by exact coefficient scaling. These derived results do not add duplicate on-screen rows. Arbitrary functions and unidentified column pairs do not receive physical labels.

Fit-uncertainty documentation is now arranged in three levels by audience:

1. Understanding fit uncertainties is the short lab-facing guide: what the fitted value and +/- uncertainty mean, uncertainty choices, R-squared, reduced chi-square, and reporting. It begins from the principle that a number from a fit is not the answer; it is a measurement with assumptions attached.
2. The statistics behind the result develops residual variance, degrees of freedom, profile curvature, nuisance-parameter refitting, rank, limitations, and the reproducible 2.000 +/- 0.063 m/s example.
3. Bevington-style implementation note records the exact numerical procedure, code path, step heuristics, edge cases, and regression coverage. The inherited equation-number reference is explicitly distinguished from an edition-verified citation.