Add Indic PNC prompt-tuning review toolkit

[mohammadaaftabv]

What changed

Adds a self-contained review toolkit under examples/audio/pnc_prompt_tuning/ for optimizing and judging punctuation-and-capitalization prompts across the 12 spreadsheet-ground-truth Indic languages.

• Treats tutorials/audio/granary_v2_postprocessing/common.yaml as the sole authoritative downstream PNC normalization contract.
• Separates generator responsibility from normalizer responsibility: the model may insert only . , ? ! and permitted capitalization; raw Unicode preservation is validated before normalization.
• Applies the pinned common.yaml rules exactly and retains both raw and common-normalized candidates for audit.
• Uses common-normalized input, candidate, and reference text for judging, metrics, aggregation, and promotion decisions.
• Adds deterministic sampling, NVIDIA model routing/caching, absolute and swapped pairwise judging, calibration, metrics, promotion reporting, prompts, a Draco review launcher, and offline tests.
• Enforces output containment under /lustre/fsw/portfolios/edgeai/users/aaftabv/pnc_prompt_finetuning, including Draco's symlinked path alias without permitting physical symlink escapes.

Why

The experiment needs one reproducible target matching the existing Granary v2 postprocessing pipeline. This avoids evaluating a second native-punctuation profile or allowing generator cleanup that the authoritative normalizer already defines.

Known blocker

The pinned common.yaml currently omits the Odia Unicode block (U+0B00–U+0B7F). The contract preflight therefore deliberately reports or as incompatible and exits before any model call. This PR does not add a local workaround or a second normalization policy; common.yaml itself must be updated and repinned before a 12-language run can proceed.

Validation

• 45 passed in the offline pytest suite.
• Ruff: all checks passed.
• bash -n passed for scripts/run_draco_review.sh.
• Secret-pattern scan found no credentials.
• Draco transfer checksum matched the local archive.
• Draco CLI import and launcher syntax checks passed.
• Draco contract report: 40 rules, SHA-256 c7c7a0b0033831551e986b2135324af409516a792c52d251d0d486d364201f81, incompatible language or, intentional exit code 2.
• No NVIDIA model/API calls were made.

Review scope

All changes are new files under examples/audio/pnc_prompt_tuning/. Existing Curator source files, common.yaml, and the production pnc_prompt.md are unchanged.

[github-actions]

This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.

[github-actions]

This PR was closed because it has been inactive for 7 days since being marked as stale.