Struct ProvenancePatch 

pub struct ProvenancePatch {
    pub license: Option<String>,
    pub data_source: Option<String>,
    pub data_license: Option<String>,
    pub hf_architecture: Option<String>,
    pub hf_model_type: Option<String>,
    pub architecture: Option<String>,
    pub tokenizer_vocab: Option<Vec<String>>,
    pub tokenizer_merges: Option<Vec<String>>,
    pub tokenizer_model_type: Option<String>,
}

In-place field patches. None means “leave unchanged”; Some("") is a legitimate explicit clear (not currently contract-approved but kept distinct from None so callers can express intent).

PMAT-690 P0-K extension (2026-05-17): hf_architecture and hf_model_type were added so pre-P0-K APRs can be patched in place without re-import. The §86 SPEC amendment surfaced this: P2-E’s epoch-49 checkpoint (val_loss=4.62, the best MODEL-2 result on record) has architecture=“LlamaForCausalLM” (the P0-H fallback) and hf_architecture=null because its init APR pre-dates P0-K. Without in-place stamping, the 50 P2-E checkpoints (~125 GB) are unusable as --init for resume training because apr pretrain reads the (wrong) architecture stamp and rejects the load. Stamping the correct hf_architecture + a corrected architecture family slug salvages the entire run without a 53-min retrain.

Fields

license: Option<String>

SPDX license identifier to stamp into the metadata.

data_source: Option<String>

Training-data source (dataset identifier or “teacher-only”).

data_license: Option<String>

SPDX license for data_source.

hf_architecture: Option<String>

HuggingFace class name from config.json::architectures[0] (e.g., “Qwen2ForCausalLM”). PMAT-690 P0-K extension.

hf_model_type: Option<String>

HuggingFace config.json::model_type (e.g., “qwen2”). PMAT-690 P0-K extension.

architecture: Option<String>

Lowercase architecture family slug (e.g., “qwen2”, “llama”). PMAT-690 P0-K extension. Distinct from hf_architecture (which is the HF class name like “Qwen2ForCausalLM”). This is the field that apr pretrain --init reads for arch dispatch, so patching this is what makes a pre-P0-K checkpoint resumable.

tokenizer_vocab: Option<Vec<String>>

Tokenizer vocabulary (token strings indexed by token-id). When Some, the stamp embeds these strings into metadata.custom["tokenizer.vocabulary"] (as a JSON array) AND sets the HAS_VOCAB header flag — making the resulting APR self-contained for apr run inference (which rejects APRs without an embedded tokenizer per PMAT-172).

PMAT-690 P3-C-prep follow-up (2026-05-17, defect 1 from publish-readiness preflight on P2-E ep49): pre-P0-K APRs lack embedded tokenizers because the training init didn’t have one. Without this stamp extension, the §86 salvage produces a 6.0 GB HF-publish-ready directory that fails the headline apr run smoke test.

tokenizer_merges: Option<Vec<String>>

BPE merge rules (e.g., ["Ä t", "i n", ...]). When Some, embedded into metadata.custom["tokenizer.merges"].

tokenizer_model_type: Option<String>

Tokenizer model type (e.g., “BPE”, “Unigram”). Optional metadata for apr inspect to surface.

Implementations

impl ProvenancePatch

pub fn has_any(&self) -> bool

true iff at least one field would change. Guards against a no-op rewrite producing a pointless new file.

Trait Implementations

impl Clone for ProvenancePatch

fn clone(&self) -> ProvenancePatch

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for ProvenancePatch

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for ProvenancePatch

fn default() -> ProvenancePatch

Returns the “default value” for a type.