[Docs] Close late code-freeze docs and release-note gaps

[lbliii]

Context

Child of #2118. The original epic used the v1.2.0 publication timestamp as its cutoff. Sara's follow-up inventory identified release-relevant changes merged after the 26.04 code freeze but before that publication timestamp. These commits are reachable from v1.2.0..main and need a final documentation/release disposition.

This issue bundles the smaller gaps that do not justify independent feature pages.

Documentation update

Semantic deduplication

Document #1927 in the semantic-deduplication guide:

• KMeansStage.fit_data_fraction
• SemanticDeduplicationWorkflow.fit_data_fraction
• TextSemanticDeduplicationWorkflow.kmeans_fit_data_fraction
• Centroid cache_path behavior
• Two-pass memory/IO tradeoffs, file-level sampling, validation, and output artifact

26.06 release-note coverage

Add or verify concise entries for:

• #1746 — repository Getting Started agent skill
• #1762 — JsonlWriter preserves UTF-8 by default
• #1763 — distinct Nemotron-CC stage names in metrics and diagnostics
• #1774 — interleaved image-text getting-started workflow
• #1870 — interactive DNS Challenge Read Speech tutorial
• #1880 — corrected URL matching for URL-ratio/repetition filters
• #1888 — lazy text-classifier imports and lower cold-import cost
• #1890 — lazy OpenAI client initialization
• #1927 — memory-efficient KMeans fitting and optional centroid caching
• #1895 / #1957 — final CUDA, Ray, PyTorch, vLLM, HAProxy, and ai-dynamo dependency/runtime changes

Existing workstream attribution

Update the existing children and release notes rather than create duplicate feature pages:

• #1427 belongs to the Nemotron-CLIMB/Megatron output workstream
• #1679 belongs to the audio-tagging workstream
• #1844 belongs to the translation workstream
• #1855 belongs to the 26.06 Python migration

Acceptance criteria

• Semantic-deduplication docs explain all #1927 controls and their memory/IO consequences
• Every PR listed above has an explicit release-note or existing-workstream disposition
• Release dependency versions reflect final main, not an intermediate PR description
• Existing interleaved and Read Speech pages are linked rather than duplicated
• #1427, #1679, #1844, and #1855 are recorded on their existing child issues
• The parent epic records the supplemental code-freeze audit
• Fern validation and link checks pass

Related PRs: #1427, #1679, #1746, #1762, #1763, #1774, #1844, #1855, #1870, #1880, #1888, #1890, #1895, #1927, #1957

[lbliii]

Implementation coverage is ready for review:

• [codex] Document memory-efficient semantic dedup fitting #2149 documents the Add fit_data_fraction and cache_path to KMeansStage #1927 semantic-deduplication controls, file-level sampling, strict validation, two-pass memory and I/O tradeoffs, and direct KMeansStage.cache_path centroid artifacts.
• Existing release-notes PR [codex] publish 26.06 release notes and migration checklist #2143 was rebased and updated with every supplemental release-note item from this issue, exact dependency versions from current main, links to the existing interleaved and Read Speech guides, and the now-released Pipeline resumability via source-level counter checkpointing #2063 resumability behavior.
• [Tutorial] Merge multiple file prefixes generated by MegatronTokenizerWriter #1427, Add generic audio processing pipeline for ASR and TTS data preparation #1679, Add translation #1844, and Bump python requirements to 3.11 min and 3.13 max #1855 are already attributed in [Docs] Add a Nemotron-CLIMB data-curation recipe #2119, [Docs] Add an audio tagging, quality-metrics, normalization, and LLM PnC guide #2123, [Docs] Complete the experimental translation configuration reference #2126, and [Docs] Publish 26.06 release notes and migration checklist #2120 respectively.
• Parent [Docs epic] Close post-v1.2.0 documentation gaps for 26.06 #2118 already records the supplemental code-freeze audit and disposition.

Validation: fern check reports 0 errors on both branches. The release-note branch link audit has 9 planned sibling-PR links plus the 22 pre-existing repository errors; #2149 introduces no new link errors.