docs: add "How training works" page

[divyasinghds]

Summary

• Adds tools-help/how-training-works.mdx — a transparency page that walks through what the tracebloc client does to your data and model in each of the nine supported use cases.
• Each accordion covers: input format, preprocessing (incl. split strategy), training/validation step, default loss/optimizer, cycle-level metrics with the underlying library calls (sklearn, torchmetrics, lifelines), and inference output.
• Closes with a 6-step "reproduce locally" checklist so an evaluating user can run the same pipeline on the same data and compare metrics number-for-number.
• Wired into the Tools & Help group in docs.json.

Replaces #25 (renamed from "pipeline reference" → "How training works"; rebased onto develop).

Test plan

• mint dev renders the page and the new nav entry appears under Tools & Help
• All nine accordions expand correctly
• mint broken-links passes
• Spot-check by an owner of core/metrics/ and core/domains/ in tracebloc-client that the per-use-case metric lists and loss formulas are accurate (notably segmentation boundary metrics and the Cox loss description)

🤖 Generated with Claude Code

---

Note

Low Risk
Low risk documentation-only changes; main risk is incorrect or out-of-date training/metric descriptions misleading users, not runtime behavior.

Overview
Adds a new join-use-case/how-training-works.mdx doc that подробно describes the platform’s training/inference pipeline and per-use-case preprocessing, loss/optimizer behavior, metrics, and inference outputs, plus guidance for reproducing runs locally.

Wires the new page into the Join a Use Case navigation in docs.json, explicitly ignores .github/ in .mintignore to prevent Mintlify dev server MDX parsing issues, and cleans up a minor formatting/encoding issue at the start of evals.json.

^{Reviewed by Cursor Bugbot for commit 6eb7094. Bugbot is set up for automated code reviews on this repo. Configure here.}

[LukasWodka]

👋 Heads-up — Code review queue is at 12 / 8

Above the WIP limit. The team convention is to review existing PRs before opening new work.

Open PRs currently in Code review (oldest first):

• client#108 — chore: sync main → develop after misrouted docs PRs · author: @saadqbal · no reviewer assigned
• docs#14 — docs: migrate join-use-case API examples to snake_case (0.7.0) · author: @divyasinghds · no reviewer assigned
• docs#4 — docs: switch install to extras syntax for tracebloc_package 0.6.33 · author: @LukasWodka · reviewer: @saadqbal
• documentation-lagecy-#7 — docs: migrate start-training and hyperparameters to snake_case API (0.7.0) · author: @aptracebloc · reviewer: @LukasWodka
• model-zoo#59 — fix: pin revision on every from_pretrained() call · author: @aptracebloc · no reviewer assigned
• start-training#10 — ci: add WIP-limit-check caller workflow · author: @LukasWodka · no reviewer assigned
• start-training#6 — chore: add PR template + customer bump + stale auto-close · author: @LukasWodka · no reviewer assigned
• start-training#9 — ci: add kanban closure-routing caller workflow · author: @LukasWodka · reviewer: @saadqbal
• tracebloc-py-package#117 — TEST: trigger WIP-limit-check (do not merge — will close shortly) · author: @saadqbal · no reviewer assigned
• tracebloc-py-package#123 — chore: tighten huggingface extra pins to compatible major-version ranges · author: @aptracebloc · no reviewer assigned

Pull from review before opening new work. (This is a nudge from the kanban WIP check, not a block.)