Repair approximate slide alignment and simplify regional correction

[Enksodsoon]

Linked panes could remain uninitialized when their reference slide was hidden, and active refinement blocked manual correction. This repairs supported approximate navigation, rejects invalid/stale anchor paths, and adds immutable region-scoped one/two-point Preview/Save/Cancel correction while preserving original slide pixels and saved corrections through background publication.

The workflow is Slides → Sync → Adjust region, with two panes by default, support for four, and engine/display details in Advanced. Existing sharing permissions, themes, keyboard controls, reset and maximize remain available.

Nine registration recipes cover native, VALIS, wsireg, HISAlign, separate DeeperHistReg classical/learned configurations and three hybrids. Optional/research-restricted components remain separate from production admission. Benchmarks retain input/settings/version/resource provenance, rejection and unsupported denominators, cold/fresh-process history, same-process repeats, per-tissue/stain results and failed runs.

Validation:

• Implementation source 9eab3edeeedded8b2fca9c0521c88cdec388e0cc: all eight jobs/checks passed in run37808320998; backend2114/115, capacity contracts508, web759, browser289/3, PostgreSQL157/1, fullstack35/one existing opt-in skip, plus QEMU ARM native/session cases.
• Final warm campaign:144 pair/recipe attempts,288 planned ordinals;262 proven executed,13 unknown,13 not executed. Startup/terminal source/runtime/inputs matched; all owned workers ended. Expanded development retains488 attempts over122 deduplicated ordered pairs.
• Final publication 747b268206991eb3ef62d2f416cfe9d442cd457b is documentation only; implementation bytes remain identical to frozen source9eab. The initial publication1e613 passed all eight CI jobs in run37843677237. Final additive stage-report hashes, finite JSON, all31 source-module bindings and full repository privacy audit passed; new-head CI runs separately. Historical CI failures remain failed.

No method is a qualified Fast/Accurate winner. Native has360/947 supported landmarks (38.01%); VALIS/wsireg each accepted one negative challenge on both calls. The original warm scorer leaves96 stages unmeasured. Separate producer-bound full-JPEG analysis measures24 saved stages and retains72 unmeasured, without changing original reports/guards or rerunning registration. Common-landmark improvements and regressions are retained; no overall hybrid superiority is claimed. Earlier12 cold stage observations remain separate. Higher observed kernel memory peaks, missing execution receipts and unmeasured cache/queue scopes are retained without invented causes or ceiling-compliance claims.

Calibrated disjoint held-out evaluation, independent anatomy review, human correction effort, native/Windows ARM hardware and default heavy-child cold startup remain unverified. No merge, deployment or production activation is included.

Full evidence: docs/alignment-results/implementation-2026-10-08/README.md and docs/alignment-results/public-warm-process-2026-10-08/VERIFICATION.md.