[Docs] Publish 26.07 release notes and migration checklist #2120

Context

Child of #2118. The current main release-notes page still presents 26.04. This task is the release-level safety net for every user-visible PR merged after v1.2.0.

Required coverage

New capabilities

• Nemotron-CLIMB recipe: Nemotron-CLIMB data curation recipe #1727, Minor updates to CLIMB tutorial #2072
• Audio tagging/quality/PnC stages: Add filtration and PnCWithLLM stages for audio tagging pipeline #1863
• Long-form audio cutting for ALM pretraining: Add long-form audio cutting pipeline for ALM pretraining #1898
• Nemotron OCR SDG: Nemotron OCR SDG Pipeline #1899
• Video caption quality evaluation: caption quality evaluation #1980
• Pipeline checkpoint/resume: Pipeline resumability via source-level counter checkpointing #2033
• Translation controls: Expose translation control knobs #2038
• Nemotron-Parse metrics/tuning: feat(nemotron-parse): richer inference metrics, engine_kwargs passthrough, fanout stage spec #2054
• Stage scheduling and with_() backend overrides: Update Ray Data Backend to use compute strategy + Allow min/max/initial workers to be specified #2077, feat: support stage config overrides via with_ #2086

Breaking or migration-sensitive behavior

• Deterministic framework-managed task_id; remove user-set examples: Unify task identity: collapse _uuid into deterministic task_id #2036, Strip user-set task_id from tutorials & getting-started script #2058
• EmptyTask singleton changed to the EmptyTask() class and SentinelTask introduced: Refactor empty/sentinel tasks: EmptyTask class + SentinelTask base #2062, Fix EmptyTask references in benchmark and tutorials #2070
• model_does_preprocess / --captioning-model-does-preprocess removed: [video] Remove model_does_preprocess #2109
• ASR dataloader num_workers renamed to dataloader_num_workers; num_workers reserved for stage worker counts: feat: support stage config overrides via with_ #2086
• nemo_text_processing unavailable from the audio extra on aarch64: fix: gate nemo_text_processing to x86_64 to unblock aarch64 build #2049
• Direct flash-attn dependency removed from video_cuda12: Remove FlashAttn Deps #2107

Improvements and fixes worth surfacing

• Filetype-aware dedup input-extension defaults: fix: default workflow input extensions by filetype #2045
• FLEURS Hugging Face caching/retry behavior: fix(audio): download FLEURS via hf_hub_download to avoid HTTP 429 #2057
• ArXiv extraction now uses a system temp directory: fix: extract arxiv archives outside download dir #2055
• Dynamo multimodal/hybrid-KV/subprocess fixes: fix(dynamo): multimodal chat processor, hybrid KV cache compat, subprocess env passthrough #2053
• NDD tutorial worker override and InferenceServer hang note: Fix NDD tutorial worker override #2104, Add note about InferenceServer hang in ndd_data_generation_example tutorial #2114
• Port scan includes 65535: fix(utils): include port 65535 in get_free_port() scan range #2024
• Video CPU-stage fusion/performance improvement: [Curator Video] Enable Ray Data stage fusion via cpus=1.0 in CPU video stages  #2110

Supplemental code-freeze coverage

Sara's follow-up audit identified release-relevant changes merged after the 26.04 code freeze but before the v1.2.0 publication timestamp.

New capabilities and tutorials

• Typed InferenceServer configuration and Dynamo aggregated/disaggregated backends: refactor: extract Ray Serve backend boundary - inference (1/5) #1813, feat!: Typed Serve Config + Dyamo config stub - inference (2/5) #1820, feat: Placement Group based subprocess manager - inference (3/5) #1833, feat: (experimental) Dynamo aggregated backend - inference (4/5) #1834, feat: (experimental) Dynamo Disaggregated serving + validators + router polish - inference (5/5) #1861
• Qwen 3 and explicit qwen2.5 video-caption variants: Add qwen3 #1827
• Nemotron 3 Nano Omni video captioning: Add Nemotron Omni Model for Video caption pipeline #1883
• Interleaved image-text getting-started workflow: Tutorials: Add interleaved getting-started tutorial #1774
• Interactive DNS Challenge Read Speech tutorial: feat(tutorials/readspeech): add interactive Jupyter notebook tutorial #1870
• Repository Getting Started agent skill: Add Skill to Nemo Curator for "Getting Started" #1746
• Memory-efficient semantic KMeans fitting and centroid caching: Add fit_data_fraction and cache_path to KMeansStage #1927

Additional breaking or dependency-sensitive behavior

• InferenceModelConfig replaced by typed Ray Serve/Dynamo model and server configs: feat!: Typed Serve Config + Dyamo config stub - inference (2/5) #1820
• Generic video-caption qwen variant replaced by qwen2.5 / qwen3: Add qwen3 #1827
• Python 3.11 minimum and Python 3.13 support: Bump python requirements to 3.11 min and 3.13 max #1855
• CUDA 12.9.1, Ray 2.55.1, PyTorch 2.10, vLLM 0.18.x, HAProxy ingress, and ai-dynamo 1.1.0 runtime changes: Bump Ray 2.55.1 / CUDA 12.9.1 / torch 2.10 / vllm 0.18; HAProxy ingress #1895, chore: Update dynamo to 1.1.0 #1957

Additional improvements and fixes

• JsonlWriter preserves UTF-8 by default: fix: preserve utf-8 characters in JSONL output #1762
• Nemotron-CC stages expose distinct names in metrics and diagnostics: fix: give nemotron-cc stages distinct names #1763
• URL matching no longer consumes adjacent HTML/punctuation in heuristic filters: fix(filters): escape - in regex_url character class (#1601) #1880
• Text classifiers use lazy package imports to reduce cold-import cost: Lazy import for text module #1888
• OpenAI clients initialize lazily when queried: Lazy init for openai client #1890

Existing feature entries to expand

• Megatron tokenizer-output merging: [Tutorial] Merge multiple file prefixes generated by MegatronTokenizerWriter #1427
• Foundational audio tagging pipeline: Add generic audio processing pipeline for ASR and TTS data preparation #1679
• Initial translation package and backends: Add translation #1844

Detailed docs work is tracked by #2144, #2145, and #2146.

Target files

• fern/versions/main/pages/about/release-notes/index.mdx
• Add or update a migration page if the release-notes page would become unwieldy
• Update Fern navigation/redirects if a new migration page is introduced

Acceptance criteria

• Page title, warning callouts, and release framing say 26.07, with no stale 26.04 framing
• Each PR above appears under a clear New / Improvement / Fix / Breaking / Dependency heading
• Every breaking item has before/after code or a concrete migration instruction
• Feature bullets link to the detailed pages created by sibling issues under [Docs epic] Close post-v1.2.0 documentation gaps for 26.07 #2118
• Dependency/platform notes distinguish x86_64 and aarch64 behavior
• Claims and parameter names are checked against merged code, not only PR descriptions
• Fern validation and link checks pass