@samesake/server@2.6.0

2.4.1

Patch Changes

• 87a8d9c: Use workspace:^ for inter-package dependencies. In dev this always resolves to the local
  workspace package; at publish bun publish rewrites it to a real ^<version> (verified via
  bun pm pack). Replaces the previous loose ^2.0.0 ranges that could silently resolve to a stale
  published version (the bug that left apps/playground pinned to ^1.3.0).

@samesake/server changelog

0.5.4 — 2026-05-20

Framework owns the schema contract; consumer apps own the prompt body.

Restructured parseService so the framework's default no longer carries
domain-specific content. Reviewer feedback: the 0.5.3 default had
Sri Lankan SME examples, currency conventions, and Sinhala/Tamil
language rules — all project-specific content that mis-steers any
other consumer (healthcare, global retail, etc.).

Changes

• PRODUCT_PARSE_SCHEMA_CONTRACT (new, exported): a small block that
  defines only the schema contract and the cross-script invariant.
  parseService ALWAYS prepends this to the final prompt, whether
  the consumer overrides or not. Consumers do not need to restate
  the contract in their override.
• DEFAULT_PRODUCT_PARSE_BODY (new): the minimal generic role-block
  used when entity.parse.instructions is not provided. No domain
  content, no examples — just "you parse one product name, faithful
  extraction is your only job".
• DEFAULT_PRODUCT_PARSE_INSTRUCTIONS retained as a deprecated alias
  to DEFAULT_PRODUCT_PARSE_BODY for 0.4.x callers. Will be removed
  in 0.7.x.
• parseService.parseProductName composes the final prompt as
  PRODUCT_PARSE_SCHEMA_CONTRACT + "\n\n" + (instructions ?? DEFAULT_PRODUCT_PARSE_BODY).

Migration

The parse-cache key includes a hash of the final composed prompt, so
this change invalidates all cached parse results automatically.

Consumers whose 0.5.3 override was a full prompt should split it into:

1. Their domain content (role + examples + extraction rules) → keep
   in their entity's parse.instructions.
2. Schema contract — DELETE from the consumer override; framework now
   provides it.

For demo consumers: the Sri Lankan SME prompt previously living in
the framework moved to the consumer's entity config as the
parse.instructions on the stockbook_item entity. Same content
(OCR digit-letter rule, brand-position rule, 6 examples including the
Sinhala "කිස්ට් ඇපල් නෙක්ටා 5OOml" case) — different location.

0.5.3 — 2026-05-20

Strengthened default product-parse prompt — addresses the residual
"4OOg" OCR digit-letter miss documented in
docs/baselines/2026-05-20-cross-script-AFTER.md §2.

DEFAULT_PRODUCT_PARSE_INSTRUCTIONS restructured per Anthropic /
GPT-5 / Vercel AI SDK prompt-engineering guidance:

• XML-tagged sections: <role>, <rules> with id'd rule blocks,
  <examples> with 6 curated input→output pairs covering the
  measured failure modes, <output_format>.
• New rule ocr_digit_letter: explicit instruction to normalise
  letter-O / letter-l confusions inside size tokens (the
  load-bearing change for the "Anchor full creme milk pwdr 4OOg"
  case that was the only miss after 0.5.2).
• New rule preserve_language distinguishes 'item' (original
  script) from 'item_canonical' (lowercase Latin), removing
  earlier ambiguity that surfaced as inconsistent cross-script
  parse output.
• New example with Sinhala 'කිස්ට් ඇපල් නෙක්ටා 5OOml' input,
  demonstrating both the original-script preservation in 'item'
  and the OCR normalisation in size_value.

Migration

The parse-cache key includes a hash of the instructions string, so
this change invalidates all cached parse results automatically.
Existing rows in the per-project entity__match tables retain
their stored parsed columns; they will not be re-parsed unless the
caller re-upserts. For demos and design-partner deploys, wipe and
re-seed; for production, re-upsert affected rows when convenient.

No DDL change. No schema-gen change.

0.5.2 — 2026-05-20

Fail-loud parse for parse-shape entities.

Discovered while testing the stockbook inventory matcher: upsert.ts
caught parseProductName failures and silently stored rows with NULL
brand/item/size_value/size_unit. Those rows were then un-matchable by
the brand_gate, size_unit_gate, and item-cosine channels downstream.
The original symptom was 2/10 seeded stockbook items having empty
parse data (Gemini was rate-limited during seed; the catch swallowed
the error).

Changes

• parseService.parseProductName now retries on transient
  user-parse failures with exponential backoff (0.3s, 1s, 3s). After
  all retries exhausted, throws with the full error chain.
• upsert.ts no longer catches parse failures — the error propagates
  to the caller (seed script, API route, etc.). The caller decides
  whether to retry the upsert or surface the failure to the user.
  This prevents the silent-NULL-row class of bug.

Migration

No DDL change. Existing 0.5.x deployments can adopt directly. Rows
that were upserted under 0.5.0 / 0.5.1 with NULL parse data are
still un-matchable until they are re-upserted (parse step then runs
with retry); consumers should re-upsert any rows whose brand_normalised
is NULL on a parse-shape entity if matching is required.

0.5.1 — 2026-05-20

Production fix for Tamil↔Latin and Sinhala↔Latin same-name matching.

Background

Adversarial Sinhala+Tamil customer-list testing (see
docs/baselines/2026-05-20-cross-script-baseline.md in the
consuming @samesake/core repo) found two production defects on the
Day-1-import flow for Sri Lankan SME customer books:

• Tamil ↔ Latin same-name pairs (e.g. Arun Sillarai ↔ அருண் சில்லரை)
  produced divergent phonetic keys (RNSLR vs RNCLR) because
  Tamil ச was mapped to category C. In Sri Lankan / modern
  Tamil, ச at word-start is the s sound — the per-script map
  was factually wrong.
• Even when phonetic keys converged (e.g. Anuja Wiwarana ↔ අනූජ විවරණ both → NCVRN), the trigram channel returned 0 because
  the two scripts share no character n-grams in their original
  form. With cosine alone carrying ~85% of the combined-score
  weight, cross-script same-name pairs sat at the 0.78 suggest
  threshold instead of clearly above it.

Changes

samesake_phonetic (system DDL)

• Tamil ச → S (was C). Aligns with how Latin s already maps.
• Tamil ஜ → C explicitly (was bundled with ச in 'சஜ' → 'CC').
  Preserves the j-class mapping to align with Latin j.

Generated match_<kind> SQL (people-shape)

• Trigram channel is now GREATEST(similarity(query.norm, candidate.name_normalised), similarity(query.phon, candidate.phon_hash)). Intra-script pairs still use the richer
  normalised-text similarity (no behavior change). Cross-script
  pairs gain a trigram bridge via their phonetic signatures — for
  identical phonetic keys, trigram ≈ 1.0 instead of 0.

The parse-shape (asset / stockbook) match function is unchanged —
its trigram channel uses the same name_normalised form, but
parse-shape entities don't rely on cross-script phonetic equivalence
in the production workloads we have today. (If they do later, the
same change applies there.)

Migration / cache implications

The samesake_phonetic change means stored name_phon values
computed by 0.4.x are stale for any row whose name contains Tamil
ச. Existing 0.4.x deployments need to re-upsert affected rows to
recompute name_phon before the matcher returns correct results.
The embedding cache (samesake_embed_cache) is unaffected —
embeddings did not change.

For new deployments: just run matcher.apply() / matcher.upsertOne()
on the new version. The DDL is idempotent; CREATE OR REPLACE FUNCTION swaps in the new phonetic logic.

Acceptance evidence

See docs/baselines/2026-05-20-cross-script-baseline.md §5 for the
required pass/fail per row on the adversarial Sinhala + Tamil blobs.
A focused phonetic smoke lives at
scripts/cross-script-smoke.ts in the consuming repo.

0.4.3 — 2026-05-19

Per-entity channel weights honored in generated SQL.

0.4.2 — 2026-05-19

Per-scope thresholds wildcard fallback; r.candidates strictly above suggest.

0.4.1 — 2026-05-19

Dedup-function name-field bug; honors entity nameField (not hardcoded a.name).

@samesake/cli@2.6.0

@samesake/cli

2.0.1

Patch Changes

• 87a8d9c: Use workspace:^ for inter-package dependencies. In dev this always resolves to the local
  workspace package; at publish bun publish rewrites it to a real ^<version> (verified via
  bun pm pack). Replaces the previous loose ^2.0.0 ranges that could silently resolve to a stale
  published version (the bug that left apps/playground pinned to ^1.3.0).
• Updated dependencies [87a8d9c]
  • @samesake/server@2.4.1