root b216b7e5b6 fix the other 4: close all OPEN-list items in one wave
Substantial wave addressing all 4 prior OPEN items. Three closed in
full, one partially (the speculative half deliberately deferred).

OPEN #1 — Periodic fresh→main index merge (FULL):
- POST /v1/vectors/index/{src}/merge with {dest, clear_source}
- Idempotent on re-runs (existing-in-dest items skipped)
- internal/vectord/index.go: new Index.IDs() snapshot method +
  i.ids tracker field as canonical ID set, independent of meta
  map's nil-vs-{} sparseness (was a real bug — IDs() backed by meta
  alone missed items added with nil metadata)
- 4 cmd-level integration tests (happy path drain+clear, dim
  mismatch, dest not found, self-merge rejection) + 1 unit test
- DecodeIndex backward-compat: old envelopes restore i.ids from
  meta keys (best effort; new items going forward use the tracker)

OPEN #2 — Distillation SFT export (SUBSTRATE):
- internal/distillation/sft_export.go ports the load-bearing half:
  IsSftNever predicate + ListScoredRunFiles (data/scored-runs/YYYY/
  MM/DD walk) + LoadScoredRunsFromFile + partial ExportSft.
- Synthesis (instruction/input/response generation) deferred to a
  separate wave — too big for this session, but the substrate
  makes the next wave a port-not-design exercise.
- TestSftNever_PinsExpectedSet locks the contamination firewall
  set: if a future commit adds/removes from SftNever, this test
  fails — forcing the change through review.
- 5 new tests; firewall fires end-to-end through the partial port.

OPEN #3 — Distribution drift via PSI (FULL):
- internal/drift/drift.go: ComputeDistributionDrift via Population
  Stability Index. Standard finance/risk metric, well-defined
  verdict tiers (stable < 0.10, minor 0.10–0.25, major ≥ 0.25).
- Equal-width bucketing over combined min/max so neither dist
  falls outside; epsilon-clamping for empty buckets so log doesn't
  blow up. Per-bucket breakdown for drilldown.
- Pairs with the existing ComputeScorerDrift: scorer drift is
  categorical, distribution drift is continuous. Different shapes,
  same package.
- 7 new tests covering identical-is-stable, hard-shift-is-major,
  moderate-detected-not-stable, empty-inputs-safe, all-identical-
  safe, bucket-counts-conserved, num-buckets-clamping.

OPEN #4 — Ops nice-to-haves (PARTIAL — wall-clock done, others
deferred):
- (a) Real-time wall-clock for stress harness: per-phase elapsed
  time logged to stdout as it runs (`[stress] phase NAME starting
  (T+12.3s)` + `[stress] phase NAME done — 8.5s (T+20.8s)`).
  Output.PhaseTimings + Output.TotalElapsedMs in JSON.
- (b) chatd fixture-mode S3 mock + (c) liberal-paraphrase
  calibration: not actioned — no fired trigger, would be
  speculative. Documented as deferred-until-need rather than
  ignored. Per the project's discipline ("don't add features
  beyond what the task requires").

OPEN list now empty / steady-state. Future items will land as
production triggers fire.

Build + vet + tests green; 18 new tests across the 4 closures.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-04-30 23:42:11 -05:00

215 lines
7.1 KiB
Go

package distillation
// SFT (Supervised Fine-Tuning) export pipeline. Closes the SUBSTRATE
// half of OPEN #2 — types, contamination firewall, file-listing
// helper. The actual synthesis (turning EvidenceRecord + ScoredRun
// into instruction/input/response triples) is still on the Rust
// side at scripts/distillation/export_sft.ts and will land in a
// follow-up wave.
//
// Why ship substrate without synthesis: the firewall constants and
// types are the load-bearing contamination guarantees. Once they're
// pinned in Go (with tests proving the firewall set is exactly
// {rejected, needs_human_review} and never expands), the synthesis
// port becomes a translation exercise rather than a design one.
//
// Per the project_distillation_substrate.md note: SFT_NEVER is one
// of the "what NOT to touch casually" knobs. Replicating it here in
// Go preserves the cross-runtime invariant — the contamination
// firewall fires even if the SFT export is run from the Go side.
import (
"encoding/json"
"errors"
"fmt"
"os"
"path/filepath"
"sort"
"strings"
)
// SftNever is declared in types.go (the load-bearing contamination
// firewall — pinned at the type-level so every consumer reads the
// same source of truth). IsSftNever below is the predicate
// helper; it lives here because it's specific to the SFT export
// path, not a property of the type system.
//
// IsSftNever returns true if a scored run's category is on the
// contamination firewall list. Inlinable; called per-record in the
// hot synthesis loop.
func IsSftNever(c ScoreCategory) bool {
for _, blocked := range SftNever {
if c == blocked {
return true
}
}
return false
}
// ExportSftOptions mirrors the TS shape so callers porting from
// Rust have an identity-translation surface. Root is the lakehouse
// data root (default $LH_DISTILL_ROOT or /home/profit/lakehouse).
// RecordedAt is the timestamp stamped on emitted SFT samples for
// lineage. IncludePartial toggles "emit even when evidence record
// is missing some optional fields"; DryRun skips file writes.
type ExportSftOptions struct {
Root string
RecordedAt string
IncludePartial bool
DryRun bool
}
// ExportSftResult mirrors the TS result shape exactly so a
// callable swap between sides doesn't break consumers reading the
// JSON.
type ExportSftResult struct {
ScoredFilesRead int `json:"scored_files_read"`
RecordsRead int `json:"records_read"`
RecordsExported int `json:"records_exported"`
RecordsQuarantined int `json:"records_quarantined"`
OutputPath string `json:"output_path"`
QuarantineSummary string `json:"quarantine_summary"`
}
// ListScoredRunFiles walks {root}/data/scored-runs/YYYY/MM/DD/*.jsonl
// and returns the sorted list. Empty when the dir doesn't exist
// (matches Rust behavior — caller should treat zero-files as a
// no-op, not an error).
func ListScoredRunFiles(root string) ([]string, error) {
if root == "" {
return nil, errors.New("distillation: empty root")
}
base := filepath.Join(root, "data", "scored-runs")
if _, err := os.Stat(base); os.IsNotExist(err) {
return nil, nil
} else if err != nil {
return nil, fmt.Errorf("stat %s: %w", base, err)
}
var out []string
years, err := os.ReadDir(base)
if err != nil {
return nil, fmt.Errorf("read %s: %w", base, err)
}
sortDirEntries(years)
for _, y := range years {
if !y.IsDir() {
continue
}
months, err := os.ReadDir(filepath.Join(base, y.Name()))
if err != nil {
continue
}
sortDirEntries(months)
for _, m := range months {
if !m.IsDir() {
continue
}
days, err := os.ReadDir(filepath.Join(base, y.Name(), m.Name()))
if err != nil {
continue
}
sortDirEntries(days)
for _, d := range days {
if !d.IsDir() {
continue
}
files, err := os.ReadDir(filepath.Join(base, y.Name(), m.Name(), d.Name()))
if err != nil {
continue
}
sortDirEntries(files)
for _, f := range files {
if strings.HasSuffix(f.Name(), ".jsonl") {
out = append(out, filepath.Join(base, y.Name(), m.Name(), d.Name(), f.Name()))
}
}
}
}
}
return out, nil
}
// sortDirEntries sorts dir entries by name in-place. Stable
// alphabetical so the directory walk is deterministic — important
// for the audit_baselines longitudinal signal which expects the
// same order across runs.
func sortDirEntries(entries []os.DirEntry) {
sort.Slice(entries, func(i, j int) bool { return entries[i].Name() < entries[j].Name() })
}
// LoadScoredRunsFromFile reads a JSONL of ScoredRun records.
// Returns the slice + the count of malformed lines (skipped).
// This is the read-half — the synthesis half (turn ScoredRun +
// EvidenceRecord into SftSample) is the not-yet-ported piece.
func LoadScoredRunsFromFile(path string) ([]ScoredRun, int, error) {
data, err := os.ReadFile(path)
if err != nil {
return nil, 0, err
}
lines := strings.Split(string(data), "\n")
out := make([]ScoredRun, 0, len(lines))
skipped := 0
for _, line := range lines {
line = strings.TrimSpace(line)
if line == "" {
continue
}
var sr ScoredRun
if err := json.Unmarshal([]byte(line), &sr); err != nil {
skipped++
continue
}
out = append(out, sr)
}
return out, skipped, nil
}
// ExportSft is the partial port. Lists scored-run files, loads
// each, applies the contamination firewall, and reports counts.
// What's NOT yet ported (deliberate, separate wave):
// - Evidence-record loading + cache (loadEvidenceByRunId).
// - synthesizeSft — the actual instruction/input/response
// synthesis logic. ~80 lines of TS in scripts/distillation/export_sft.ts.
// - Quarantine writer integration (write rejected records to
// a quarantine JSONL for operator review).
// - File output (write SFT JSONL to data/distilled/sft/).
//
// Returning a non-nil result with RecordsExported=0 is intentional
// pre-synthesis — operators calling this on the Go side will see
// the count of records that PASSED the firewall and would have
// been exported by a complete implementation. RecordsQuarantined
// reflects records BLOCKED by the firewall.
//
// Tests/contracts that synthesis port must preserve:
// - SftNever firewall fires before any other validation
// - Sort order matches Rust (file walk + record order within file)
// - Empty root dir returns zero-counts, not error
func ExportSft(opts ExportSftOptions) (ExportSftResult, error) {
res := ExportSftResult{
OutputPath: filepath.Join(opts.Root, "data", "distilled", "sft", "sft_partial.jsonl"),
QuarantineSummary: "synthesis not yet ported — see internal/distillation/sft_export.go header",
}
files, err := ListScoredRunFiles(opts.Root)
if err != nil {
return res, fmt.Errorf("list scored runs: %w", err)
}
res.ScoredFilesRead = len(files)
for _, f := range files {
runs, _, err := LoadScoredRunsFromFile(f)
if err != nil {
continue
}
res.RecordsRead += len(runs)
for _, r := range runs {
if IsSftNever(r.Category) {
res.RecordsQuarantined++
continue
}
// Synthesis would happen here. Pre-port: count as
// "would-export" for the firewall-passing records.
res.RecordsExported++
}
}
return res, nil
}