fix(interactive): show Windows path examples on Windows (client#615)

[shujaatTracebloc]

Closes tracebloc/client#615.

The bug

Step 3 of the guided ingest — "Where is your data?" — offered only POSIX examples:

Tabular   one CSV file                        e.g. ~/data/patients.csv
Images    a folder with labels.csv + images/   e.g. ~/data/xray/
Text      a folder with labels.csv + texts/     e.g. ~/data/reviews/

A Windows user has no such path. With nothing in the right shape to copy, they invent one — the
reported case entered a Python-style r"C:\Users\..." literal, which the CLI can't read because
this prompt is read literally, not shell-parsed.

The change

datasetPathExamples returns examples in the form the host OS actually uses. On Windows:

Tabular   one CSV file                        e.g. C:\Users\you\data\patients.csv
Images    a folder with labels.csv + images/   e.g. C:\Users\you\data\xray\
Text      a folder with labels.csv + texts/     e.g. C:\Users\you\data\reviews\

Two deliberate details:

• The label and padding text stays in one place — only the example is substituted — so the
  STYLE.md column alignment can't drift between two copies of the same line.
• It takes goos as a parameter rather than reading runtime.GOOS internally, so the Windows
  branch is exercised from a Linux/macOS runner where it would otherwise never be covered. The
  call site passes runtime.GOOS.

Folder examples end in a separator on both platforms: a folder example that doesn't look like a
folder invites a file path instead.

Tests

Four new, covering the things that would actually regress:

• Windows examples carry no POSIX syntax (~ or /) and folder examples end in \.
• The POSIX branch is not regressed while fixing Windows — including a bare/unknown goos.
• The three modalities stay distinguishable; identical examples would teach the user nothing.
• The call site follows the host OS.

Verified they bite: disabling the Windows branch fails the suite.

Note on the copy golden

zz-all-strings.golden now records the format strings (e.g. %s) instead of the concrete
examples, since the text is dynamic — the example values are covered by the unit tests instead.
Regenerated with TB_UPDATE_GOLDEN=1 and I reviewed the diff rather than accepting it blind:
exactly the four intended strings changed, nothing else.

Verification

go test ./... — 16 packages ok, 0 failures. gofmt -s, goimports -local, go vet,
errcheck, misspell, staticcheck -checks all,-ST1005 all clean. Builds for windows/amd64
as well as the host.

I develop on macOS, so the Windows rendering is reasoned from the format string plus
Printer.Infof (which does fmt.Sprintf) — I verified the formatting path rather than assuming
it, but the visual check on a Windows console is still worth one glance.

🤖 Generated with Claude Code

---

Note

Low Risk
Prompt-copy UX fix only; no path parsing, auth, or ingest logic changes. Risk is limited to incorrect example text on Windows.

Overview
Fixes the guided ingest "Where is your data?" step so example paths match the host OS. Windows users previously only saw POSIX ~/data/... examples and sometimes entered unreadable formats (e.g. Python r"..." literals).

Adds datasetPathExamples, parameterized by goos for testability, and wires it into the modality hints and input help text via runtime.GOOS. Windows gets C:\Users\you\data\... examples (folders ending in \); other OSes keep ~/data/....

Includes unit tests for Windows/POSIX shape, modality distinctness, and host-OS selection, and updates the strings golden to record %s placeholders for the now-dynamic examples. Bumps version to 0.10.5.

^{Reviewed by Cursor Bugbot for commit f36ede4. Bugbot is set up for automated code reviews on this repo. Configure here.}

[saadqbal]

Clean, tight fix 👍 Parameterizing datasetPathExamples on goos so the Windows branch is testable from a Linux/macOS runner is the right call, and the tests bite (verified locally — package green, vet clean). Golden switching to e.g. %s for the now-dynamic lines is correct.

One tiny non-blocker (pre-existing, not introduced here): the three e.g. markers don't quite line up in monospace — Tabular sits at col 46, Images 47, Text 48. Since the PR leans on column alignment as a design detail, might be worth a 2-space / 1-space nudge while these lines are open. Not merge-blocking.