docs: align apply-template guides with deferred-rounding pipeline

[giardiello]

Summary

Both implementer-facing docs (FDL_Apply_Template_Logic.md and FDL_Template_Implementer_Guide.md) were still describing the old "round everything to integers after scaling" behaviour. The pipeline has been reshaped around Adam's `ef57e1f6` rounding work:

• Rounding is deferred to the end of the pipeline (new Phase 9b, after crop), not applied mid-pipeline after scaling.
• Only canvas.dimensions and canvas.effective_dimensions are integerized. protection_dimensions, framing dimensions, and all anchors stay float per the "fractional pixels" model.
• effective_dims is always ceil'd -- the template's RoundStrategy only applies to canvas.dimensions.
• Rounding deltas are absorbed symmetrically into anchor positions (delta/2) so content stays centered on its pre-round center across chained template applications.
• _scaled_bounding_box and _content_translation are exported as custom attributes in their pre-round float form, so image processors (e.g. `fdl_imaging`'s warp-based pipeline) can place content at sub-pixel precision without re-quantizing.

This PR is docs-only -- no code changes. It brings both guides in sync with the current `native/core` code at `ef57e1f6`.

Changes

`FDL_Apply_Template_Logic.md`

• Pipeline ASCII diagram updated ("5. Scale (float)" + new "9b. Round & Absorb").
• Step 5 rewritten -- rounding language removed.
• New Step 9b with the full `geometry_round()` algorithm (ceil effective, round canvas, clamp inner, symmetric anchor shift, per-axis dim shrink, hierarchy re-establishment).
• Step 10 custom-attributes description clarified as pre-round float.
• "Complete Formula Reference" header now says inputs are floats.

`FDL_Template_Implementer_Guide.md`

• Section 2.1 ops table: `fdl_geometry_round()` description updated.
• Section 4.5 rewritten for the no-rounding Phase 5 semantics.
• New section 4.9 "Phase 9b: Round Integer-Typed Schema Fields" with pseudocode mirroring `geometry_round()`.
• Section 5.6 retitled "Exact Float Comparison in FIT Regime" -- explains why FIT detection stays bit-exact despite Phase 5 no longer rounding.
• Section 6.1 output-object fields annotated with float/integer status.
• Section 7 formula inputs annotated as floats.
• Section 8 summary table: phase 5 renamed, phase 9b added; "Round early" design decision replaced with "Round late, absorb into anchors"; added a 7th decision about the unrounded side-channel for imaging.

`docs/svg/02_pipeline_overview.svg`

• Phase 5 label updated.
• New Phase 9b box inserted between Phase 8 and the (now Phase 10) output step; viewBox and background extended by 76px to fit.

Test plan

Docs-only change -- no tests affected. Verification:

• Reviewer confirms the Phase 9b algorithm in both docs matches the actual C++ in `native/core/src/fdl_geometry.cpp::geometry_round()` at `ef57e1f6`.
• Reviewer confirms the custom-attribute descriptions (`_scaled_bounding_box`, `_content_translation` as pre-round floats) match what is actually written in `fdl_template.cpp::build_template_output_document()`.
• Visual sanity check of the updated `02_pipeline_overview.svg` (the `.png` sibling is stale and will need regeneration from whatever tool originally produced it).

Base

Based on `base/adam-ef57e1f6`. The docs on that base are identical to those on `origin/dev`, so this change can be re-targeted to `origin/dev` once Adam's rounding work lands upstream.

Made with Cursor