Fix recurring AI translation output bugs (fences, dup keys, truncation)

[beagandica]

What

Fixes the two recurring AI translation output bugs we kept manually cleaning up before merge in NuevoFoundation/workshops PRs (most recently #606 and #658):

1. Code-fence wrapping (~30% of files): the LLM was wrapping the entire translated file in ```markdown ... ``` or ```yaml ... ```, which breaks Hugo's frontmatter parser. PR #658 had 31/78 files affected.
2. Duplicate YAML keys: some translated answer-key files contained two hidden: true lines, which the YAML parser rejects. PR #606 had 3 such files.

Both bugs required hand-cleanup on every translation PR. After this change, the tool produces clean output, and if the LLM still does something invalid the file is skipped with a clear error instead of silently writing broken markdown.

Changes

Stronger prompt (Program.cs)

• Explicitly forbids wrapping output in any kind of code fence (markdown/md/yaml/yml).
• Forbids preamble or commentary.
• Requires translated frontmatter to have the same key set as the source (no additions, no removals, no duplicates).
• Spells out which frontmatter values may be translated (title, description, summary) vs which must be preserved byte-for-byte.

Inference settings

• Temperature: 1.0 → 0.2 (translation is deterministic, high temperature was driving the structural defects).
• MaxTokens: 1500 → 8192 (multiple workshop pages exceed 1500 tokens and were being silently truncated).

Defense-in-depth post-processing

• Strips a single layer of outer code-fence wrapping if the LLM still adds one. Conservative: only strips when the first and last non-empty lines are a matching fence pair, so legitimate code blocks inside the body are preserved.
• Also handles the rarer "fence wraps frontmatter only" pattern.
• Strips a leading BOM.
• Logs every cleanup it performs.

Validation (fail fast, never write broken output)

• Checks FinishReason — if the model hit the token limit, skip the file and exit non-zero.
• Parses translated frontmatter with YamlDotNet (WithDuplicateKeyChecking()), which throws on the exact dup-key pattern from #606.
• Verifies translated frontmatter has the same key set as source.
• Rejects translations whose body is empty when source body was not.
• Rejects translations that still start with a code fence.

Path bug

• Replaced filePath.Replace("english", newFolder) with a segment-aware path replace. The old version would corrupt any path whose filename contained the substring "english" (e.g. a workshop named "my-english-club").

Exit behavior

• Files that fail validation are skipped (no broken output written).
• The process exits with code 2 so the workflow surfaces the problem instead of silently producing bad PRs.

Dependency

• Added YamlDotNet 16.2.1 for safe frontmatter parsing.

Validation

Locally exercised the post-processor and validator against the real bug patterns we hit in PRs #606 and #658:

PASS  strip-markdown-fence-wrap            (PR #658 pattern)
PASS  strip-yaml-fence-wrap                (PR #658 pattern)
PASS  strip-partial-yaml-fence-wrap        (PR #658 pattern, rare)
PASS  passthrough-clean                    (no false positives on good output)
PASS  strip-bom
PASS  preserve-internal-code-fences        (does not strip legitimate body code blocks)
PASS  empty-response-throws
PASS  detect-key-mismatch
PASS  detect-duplicate-key                 (PR #606 pattern)
PASS  reject-residual-fence
PASS  path-replace-segment-only
PASS  reject-empty-body

Follow-up

Once this is merged, the submodule pointer in NuevoFoundation/workshops needs to be bumped via a separate PR. That PR will also need to bump the workflow's .NET SDK from 8.0.x to 10.0.x (the repo upgraded the project to .NET 10 in #6 but never bumped the submodule pointer or workflow).

---

Tracking issue: NuevoFoundation/workshops#662 (assigned to @ozhang22).