feat(opt): three efficiency levers for NL→SQL workflow

Lever 1: Join-graph via FK logics in annotations + suggest-fks command
  - TableAnnotation.foreign_keys field stores curated logical FKs (DWH has no FK constraints)
  - tht schema suggest-fks: mine from approved SQL, heuristics (time_key → dim_time),
    same-name discovery + explicit --assume flag for multi-owner PKs
  - mschema renders 【Foreign keys】 section populated; validation in merge.py
  - SKILL.md F4 now reads FKs from mschema-text, no custom data_time_key logic

Lever 2: Context-pack consolidation at kickoff (tht search pack)
  - Single embedding of question, reused for schema + evidence + solved searches
  - One command: tht search pack <question> --session <id> → retrieval_pack.md
  - Graceful degradation when Ollama/vector store unreachable (exit 0, empty sections)
  - SKILL.md F1 prescribes as first call; reduces model thinking turns via pre-retrieval

Lever 3: Phase-summary recap v2 auto-construction from session ledger
  - tht session show --json includes full decisions ledger
  - tht phase meta --json exports 'emits' (substantive decision types per phase)
  - Gate appends deterministic 【Decisioni registrate in questa fase】 section (appendLedgerSection)
  - Model authors only summary + checks; recap table comes from persisted state (exact by construction)
  - SKILL.md Disciplina 6: brief model output, gate fills the rest

Tests: 358 Python (including 10 FK + 3 pack + 1 session-ledger tests) + 111 JS gate tests, all pass.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

harness/tht/cli/phase_cmd.py

@@ -71,6 +71,7 @@ def meta_cmd(
                         "name": p.name,
                         "advance": p.advance,
                         "artifacts_out": p.artifacts_out,
+                        "emits": p.emits,
                     }
                     for p in wf.phases
                 ],

harness/tht/cli/schema_cmd.py

@@ -141,6 +141,174 @@ def check_cmd(config: Path = CONFIG_OPT) -> None:
     typer.secho("OK: nessuna annotazione orfana.", fg=typer.colors.GREEN)
 
 
+# PK con questi nomi sono identificatori generici: la regola same-name non si applica
+# (nel DWH reale `id` e' la PK di ~50 tabelle e produrrebbe migliaia di falsi positivi).
+_GENERIC_PK_NAMES = {"id", "key", "code"}
+
+
+@schema_app.command("suggest-fks")
+def suggest_fks_cmd(
+    config: Path = CONFIG_OPT,
+    from_sql: list[Path] = typer.Option(
+        None, "--from-sql",
+        help="Directory di .sql approvati da cui minare i join reali (ripetibile).",
+    ),
+    assume: list[str] = typer.Option(
+        None, "--assume",
+        help="Disambigua una PK con piu' proprietari: col=tabella_ref "
+             "(es. cod_paz=dim_patient). Ripetibile.",
+    ),
+    write: bool = typer.Option(
+        False, "--write",
+        help="Fonde i suggerimenti in annotations.yaml (aggiunge solo FK mancanti).",
+    ),
+) -> None:
+    """Suggerisce FK logiche per la curazione umana in annotations.yaml.
+
+    Tre regole, in ordine di confidenza: (1) equi-join minati dall'SQL gia'
+    approvato (--from-sql); (2) colonna `*time_key` verso la PK di dim_time;
+    (3) colonna con lo stesso nome della PK di UN'ALTRA tabella, solo se quel
+    nome ha un unico proprietario e non e' generico (id/key/code) — salvo
+    disambiguazione esplicita con --assume.
+    """
+    import yaml as _yaml
+
+    from tht.mschema.fkmine import mine_join_pairs
+    from tht.mschema.models import Annotations, ForeignKey, PhysicalSchema, TableAnnotation
+
+    cfg = _load_config_or_exit(config)
+    phys_file = physical_path(cfg)
+    if not phys_file.exists():
+        typer.secho(
+            f"ERRORE: {phys_file} non trovato. Esegui prima `tht schema introspect`.",
+            fg=typer.colors.RED, err=True,
+        )
+        raise typer.Exit(code=1)
+    physical = PhysicalSchema.from_yaml(phys_file)
+    ann_path = annotations_path(cfg)
+    annotations = Annotations.from_yaml(ann_path)
+
+    assumed: dict[str, str] = {}
+    for a in assume or []:
+        col, _, ref = a.partition("=")
+        if not ref or ref not in physical.tables:
+            typer.secho(
+                f"ERRORE: --assume '{a}' non valido (atteso col=tabella nel catalogo).",
+                fg=typer.colors.RED, err=True,
+            )
+            raise typer.Exit(code=1)
+        assumed[col] = ref
+
+    def _single_pk(table) -> str | None:
+        pks = [c for c, col in table.columns.items() if col.pk]
+        return pks[0] if len(pks) == 1 else None
+
+    pk_owners: dict[str, list[str]] = {}
+    for tname, table in physical.tables.items():
+        pk = _single_pk(table)
+        if pk:
+            pk_owners.setdefault(pk, []).append(tname)
+
+    dim_time_pk = None
+    if "dim_time" in physical.tables:
+        dim_time_pk = _single_pk(physical.tables["dim_time"])
+
+    def _known(tname: str) -> set:
+        keys = set()
+        for fk in physical.tables[tname].foreign_keys:
+            keys.add((tuple(fk.columns), fk.ref_table, tuple(fk.ref_columns)))
+        ann = annotations.tables.get(tname)
+        if ann:
+            for fk in ann.foreign_keys:
+                keys.add((tuple(fk.columns), fk.ref_table, tuple(fk.ref_columns)))
+        return keys
+
+    known_by_table: dict[str, set] = {t: _known(t) for t in physical.tables}
+    suggested: dict[str, list[ForeignKey]] = {}
+
+    def _add(tname: str, col: str, ref_table: str, ref_col: str) -> None:
+        key = ((col,), ref_table, (ref_col,))
+        if key in known_by_table[tname]:
+            return
+        known_by_table[tname].add(key)
+        suggested.setdefault(tname, []).append(
+            ForeignKey(columns=[col], ref_table=ref_table, ref_columns=[ref_col])
+        )
+
+    # Regola 1: join minati dall'SQL approvato.
+    n_sql_files = 0
+    mined_total = 0
+    for d in from_sql or []:
+        for sql_file in sorted(d.rglob("*.sql")):
+            n_sql_files += 1
+            pairs = mine_join_pairs(sql_file.read_text(), physical)
+            mined_total += sum(pairs.values())
+            for (src_t, src_c, ref_t, ref_c) in pairs:
+                _add(src_t, src_c, ref_t, ref_c)
+
+    # Regole 2 e 3: convenzioni di naming.
+    ambiguous_skipped: set[str] = set()
+    for tname, table in physical.tables.items():
+        for cname in table.columns:
+            if dim_time_pk and cname.endswith("time_key") and tname != "dim_time":
+                _add(tname, cname, "dim_time", dim_time_pk)
+                continue
+            if cname in assumed:
+                if assumed[cname] != tname:
+                    _add(tname, cname, assumed[cname], cname)
+                continue
+            owners = [o for o in pk_owners.get(cname, []) if o != tname]
+            if not owners or cname in _GENERIC_PK_NAMES:
+                continue
+            if len(pk_owners[cname]) > 1:
+                ambiguous_skipped.add(cname)
+                continue
+            _add(tname, cname, owners[0], cname)
+
+    if n_sql_files:
+        typer.secho(
+            f"Minati {mined_total} equi-join da {n_sql_files} file SQL.",
+            fg=typer.colors.BLUE, err=True,
+        )
+    if ambiguous_skipped:
+        typer.secho(
+            "PK ambigue saltate dalla regola same-name (piu' tabelle proprietarie): "
+            + ", ".join(sorted(ambiguous_skipped))
+            + ". Se servono, aggiungile a mano o passa --from-sql.",
+            fg=typer.colors.YELLOW, err=True,
+        )
+
+    n_fks = sum(len(v) for v in suggested.values())
+    if not suggested:
+        typer.secho("OK: nessuna FK da suggerire.", fg=typer.colors.GREEN)
+        return
+
+    if write:
+        for tname, fks in suggested.items():
+            ann = annotations.tables.setdefault(tname, TableAnnotation())
+            ann.foreign_keys.extend(fks)
+        annotations.to_yaml(ann_path)
+        typer.secho(
+            f"OK: {n_fks} FK suggerite aggiunte a {ann_path} "
+            f"({len(suggested)} tabelle). Rivedile a mano prima dell'uso.",
+            fg=typer.colors.GREEN,
+        )
+        return
+
+    payload = {
+        "tables": {
+            tname: {"foreign_keys": [fk.model_dump(exclude_defaults=True) for fk in fks]}
+            for tname, fks in suggested.items()
+        }
+    }
+    typer.echo(_yaml.safe_dump(payload, sort_keys=False, allow_unicode=True))
+    typer.secho(
+        f"{n_fks} FK candidate ({len(suggested)} tabelle). "
+        f"Usa --write per fonderle in annotations.yaml, poi curale a mano.",
+        fg=typer.colors.YELLOW,
+    )
+
+
 @schema_app.command("render")
 def render_cmd(
     config: Path = CONFIG_OPT,

harness/tht/cli/search_cmd.py

@@ -205,3 +205,156 @@ def search_cmd(
             row.append((r.content[:120] + "…") if len(r.content) > 120 else r.content)
         table.add_row(*row)
     Console().print(table)
+
+
+# Dimensioni fisse del pack (niente config: il pack deve restare piccolo perche'
+# entra nel contesto del modello in un turno solo).
+PACK_EVIDENCE_TOP = 5
+PACK_SOLVED_TOP = 3
+PACK_EXCERPT_CHARS = 400
+
+
+@search_app.command("pack")
+def pack_cmd(
+    question: str = typer.Argument(..., help="La domanda in linguaggio naturale."),
+    config: Path = CONFIG_OPT,
+    session: str = typer.Option(
+        None, "--session", help="Scrive il pack in sessions/<id>/retrieval_pack.md."
+    ),
+    json_out: bool = typer.Option(False, "--json", help="Output JSON (per Pi)."),
+) -> None:
+    """Context-pack F1: tabelle candidate + evidence + domande risolte in UNA chiamata.
+
+    Un solo embedding della domanda, riusato per le tre ricerche vettoriali.
+    Degrado gentile: se Ollama/vectordb non rispondono, le sezioni restano vuote
+    con un'avvertenza (exit 0) — la sessione prosegue con le ricerche live.
+    """
+    from sqlalchemy.exc import OperationalError
+
+    from tht.cli.vector_cmd import make_embedder, open_searcher, require_vector_cfg
+    from tht.search import combined_search, schema_tables
+    from tht.solved import SOLVED_KIND
+    from tht.vectorstore.embeddings import EmbeddingsError
+    from tht.vectorstore.rest_client import VectorRestError
+
+    cfg = _load_config_or_exit(config)
+    require_vector_cfg(cfg)
+
+    tables: list[dict] = []
+    evidence: list[dict] = []
+    solved: list[dict] = []
+    warnings: list[str] = []
+    degrade = (VectorRestError, EmbeddingsError, OperationalError)
+
+    vec = None
+    searcher = embedder = None
+    try:
+        searcher = open_searcher(cfg)
+        embedder = make_embedder(cfg.embeddings)
+        vec = embedder.embed_query(question)
+    except degrade as e:
+        warnings.append(f"retrieval non disponibile ({e}): prosegui con le ricerche live")
+
+    if vec is not None:
+        from tht.cli.schema_cmd import physical_path
+
+        descriptions: dict[str, str] = {}
+        phys_file = physical_path(cfg)
+        if phys_file.exists():
+            from tht.mschema.models import PhysicalSchema
+
+            phys = PhysicalSchema.from_yaml(phys_file)
+            descriptions = {t: tab.comment for t, tab in phys.tables.items()}
+        try:
+            cand = combined_search(
+                keyword=question, lsh_hits=None, store=searcher, embedder=embedder,
+                top=cfg.search.schema_chunk_pool, rrf_k=cfg.search.rrf_k,
+                kinds=KIND_MAP["schema"], query_vec=vec,
+            )
+            tables = [
+                {"name": n, "rrf": round(s, 6), "description": descriptions.get(n, "")}
+                for n, s in schema_tables(cand, top_tables=cfg.search.top_schema_tables)
+            ]
+        except degrade as e:
+            warnings.append(f"ricerca schema fallita ({e})")
+        try:
+            ev = combined_search(
+                keyword=question, lsh_hits=None, store=searcher, embedder=embedder,
+                top=PACK_EVIDENCE_TOP, rrf_k=cfg.search.rrf_k,
+                kinds=KIND_MAP["evidence"], query_vec=vec,
+            )
+            evidence = [
+                {"title": r.label, "status": r.status,
+                 "excerpt": r.content[:PACK_EXCERPT_CHARS]}
+                for r in ev
+            ]
+        except degrade as e:
+            warnings.append(f"ricerca evidence fallita ({e})")
+        try:
+            hits = searcher.search(vec, top_n=PACK_SOLVED_TOP, kinds=[SOLVED_KIND])
+            solved = [
+                {
+                    "session_id": h.metadata.get("session_id", h.ref),
+                    "question": h.metadata.get("question", h.content),
+                    "sql": h.metadata.get("sql", ""),
+                    "tables": h.metadata.get("tables", []),
+                    "score": round(h.similarity, 4),
+                }
+                for h in hits
+            ]
+        except degrade as e:
+            warnings.append(f"solved-search fallita ({e})")
+
+    for w in warnings:
+        typer.secho(f"ATTENZIONE: {w}", fg=typer.colors.YELLOW, err=True)
+
+    md_lines = ["# Retrieval pack", "", f"Domanda: {question}", ""]
+    md_lines += [f"## Tabelle candidate (top {len(tables)}, vettoriale sull'intera domanda)", ""]
+    if tables:
+        for i, t in enumerate(tables, 1):
+            desc = f" — {t['description']}" if t["description"] else ""
+            md_lines.append(f"{i}. **{t['name']}**{desc} (rrf {t['rrf']})")
+    else:
+        md_lines.append("_nessuna (retrieval non disponibile o nessun match)_")
+    md_lines += ["", "## Evidence rilevanti", ""]
+    if evidence:
+        for e in evidence:
+            status = f" [{e['status']}]" if e["status"] else ""
+            md_lines.append(f"- **{e['title']}**{status}: {e['excerpt']}")
+    else:
+        md_lines.append("_nessuna_")
+    md_lines += ["", "## Domande risolte simili (exemplar di riferimento, NON decisioni)", ""]
+    if solved:
+        for s in solved:
+            md_lines.append(
+                f"### {s['question']}  \n(sessione `{s['session_id']}`; "
+                f"tabelle: {', '.join(s['tables']) or '-'})"
+            )
+            if s["sql"]:
+                md_lines += ["", "```sql", s["sql"], "```", ""]
+    else:
+        md_lines.append("_nessuna_")
+    if warnings:
+        md_lines += ["", "## Avvertenze", ""] + [f"- {w}" for w in warnings]
+    md = "\n".join(md_lines) + "\n"
+
+    if session:
+        from tht.cli.session_cmd import load_session_or_exit, session_dir
+
+        load_session_or_exit(cfg, session)
+        out = session_dir(cfg, session) / "retrieval_pack.md"
+        out.write_text(md)
+        if not json_out:
+            typer.secho(
+                f"OK: retrieval pack scritto in {out} "
+                f"({len(tables)} tabelle, {len(evidence)} evidence, {len(solved)} solved).",
+                fg=typer.colors.GREEN,
+            )
+    if json_out:
+        typer.echo(json.dumps(
+            {"question": question, "tables": tables, "evidence": evidence,
+             "solved": solved, "warnings": warnings},
+            ensure_ascii=False, indent=2,
+        ))
+    elif not session:
+        typer.echo(md)

harness/tht/cli/session_cmd.py

@@ -181,6 +181,11 @@ def show_cmd(
         data = manifest.model_dump(mode="json", by_alias=True)
         data["phase"] = phase
         data["has_schema_linking"] = has_schema_linking
+        # Ledger integrale: il gate lo usa per costruire deterministicamente il
+        # recap delle decisioni nei riepiloghi di fase (v2).
+        data["decisions"] = [
+            d.model_dump(mode="json") for d in list_decisions(sdir)
+        ]
         typer.echo(json.dumps(data, ensure_ascii=False, indent=2))
         return
 

harness/tht/mschema/fkmine.py

@@ -0,0 +1,58 @@
+"""Mining dei join reali dall'SQL approvato: coppie equi-join -> FK logiche candidate.
+
+La fonte di verita' sono le query gia' validate da un umano (sql_final.sql, ctes/*.sql
+delle sessioni approvate): un equi-join ricorrente tra due tabelle del catalogo, con
+una delle due colonne PK della propria tabella, e' una FK logica ad alta confidenza.
+"""
+
+from collections import Counter
+
+import sqlglot
+from sqlglot import exp
+
+from tht.mschema.models import PhysicalSchema
+
+JoinPair = tuple[str, str, str, str]  # (src_table, src_col, ref_table, ref_col)
+
+
+def mine_join_pairs(sql_text: str, physical: PhysicalSchema) -> Counter:
+    """Estrae le coppie equi-join tra tabelle del catalogo da un testo SQL.
+
+    Ritorna un Counter {(src_table, src_col, ref_table, ref_col): occorrenze}.
+    Il lato ref e' quello la cui colonna e' PK della propria tabella; coppie in cui
+    nessuno o entrambi i lati sono PK vengono scartate (non FK-like). Alias e CTE
+    vengono risolti; i riferimenti a CTE (non nel catalogo) sono ignorati.
+    """
+    pairs: Counter = Counter()
+    try:
+        statements = sqlglot.parse(sql_text, read="postgres")
+    except sqlglot.errors.ParseError:
+        return pairs
+    for stmt in statements:
+        if stmt is None:
+            continue
+        alias_map: dict[str, str] = {}
+        for t in stmt.find_all(exp.Table):
+            alias_map[t.alias_or_name] = t.name
+        for eq in stmt.find_all(exp.EQ):
+            left, right = eq.left, eq.right
+            if not (isinstance(left, exp.Column) and isinstance(right, exp.Column)):
+                continue
+            if not (left.table and right.table):
+                continue
+            lt = alias_map.get(left.table, left.table)
+            rt = alias_map.get(right.table, right.table)
+            if lt == rt or lt not in physical.tables or rt not in physical.tables:
+                continue
+            lc, rc = left.name, right.name
+            if lc not in physical.tables[lt].columns or rc not in physical.tables[rt].columns:
+                continue
+            l_pk = physical.tables[lt].columns[lc].pk
+            r_pk = physical.tables[rt].columns[rc].pk
+            if l_pk == r_pk:  # nessuna o entrambe PK: non FK-like
+                continue
+            if r_pk:
+                pairs[(lt, lc, rt, rc)] += 1
+            else:
+                pairs[(rt, rc, lt, lc)] += 1
+    return pairs

harness/tht/mschema/merge.py

@@ -12,4 +12,14 @@ def find_orphans(physical: PhysicalSchema, annotations: Annotations) -> list[str
         for column_name in table_ann.columns:
             if column_name not in table.columns:
                 orphans.append(f"{table_name}.{column_name}")
+        for fk in table_ann.foreign_keys:
+            label = f"{table_name}.fk({','.join(fk.columns)})->{fk.ref_table}"
+            ref = physical.tables.get(fk.ref_table)
+            if ref is None:
+                orphans.append(label)
+                continue
+            missing = [c for c in fk.columns if c not in table.columns]
+            missing += [c for c in fk.ref_columns if c not in ref.columns]
+            if missing:
+                orphans.append(label)
     return orphans

harness/tht/mschema/models.py

@@ -78,6 +78,10 @@ class TableAnnotation(BaseModel):
     concepts: list[str] = []
     notes: str = ""
     columns: dict[str, ColumnAnnotation] = {}
+    # FK "logiche" curate a mano: il DWH non dichiara vincoli, quindi i join
+    # noti (es. data_time_key -> dim_time.day_key) vivono qui e vengono fusi
+    # con le FK fisiche in tutte le viste renderizzate.
+    foreign_keys: list[ForeignKey] = []
 
 
 class Annotations(_YamlModel):

harness/tht/mschema/render.py

@@ -1,11 +1,25 @@
 from typing import Any
 
 from tht.mschema.eligibility import effective_eligibility
-from tht.mschema.models import Annotations, ColumnAnnotation, PhysicalSchema
+from tht.mschema.models import Annotations, ColumnAnnotation, ForeignKey, PhysicalSchema
 
 MAX_EXAMPLES_IN_PROMPT = 5
 
 
+def table_foreign_keys(
+    physical: PhysicalSchema, annotations: Annotations, table: str
+) -> list[ForeignKey]:
+    """FK fisiche + FK logiche dalle annotations (dedup su columns/ref)."""
+    fks = list(physical.tables[table].foreign_keys)
+    ann = annotations.tables.get(table)
+    if ann:
+        seen = {(tuple(f.columns), f.ref_table, tuple(f.ref_columns)) for f in fks}
+        for fk in ann.foreign_keys:
+            if (tuple(fk.columns), fk.ref_table, tuple(fk.ref_columns)) not in seen:
+                fks.append(fk)
+    return fks
+
+
 def _ann_col(annotations: Annotations, table: str, column: str) -> ColumnAnnotation | None:
     ann = annotations.tables.get(table)
     if ann is None:
@@ -59,7 +73,7 @@ def to_mschema_text(
                 shown = ", ".join(column.examples[:MAX_EXAMPLES_IN_PROMPT])
                 lines.append(f"    -- Examples: {shown}")
         lines.append(");")
-        for fk in table.foreign_keys:
+        for fk in table_foreign_keys(physical, annotations, table_name):
             for src, dst in zip(fk.columns, fk.ref_columns):
                 fk_lines.append(f"{table_name}.{src}={fk.ref_table}.{dst}")
     lines.extend(["", "【Foreign keys】", *fk_lines])
@@ -89,7 +103,7 @@ def to_schema_dict(
             "primary_keys": [c for c in cols if table.columns[c].pk],
             "foreign_keys": [
                 {"columns": fk.columns, "ref_table": fk.ref_table, "ref_columns": fk.ref_columns}
-                for fk in table.foreign_keys
+                for fk in table_foreign_keys(physical, annotations, table_name)
             ],
         }
     return out
@@ -132,9 +146,10 @@ def to_markdown(physical: PhysicalSchema, annotations: Annotations | None = None
                 f"| {column_name} | {column.type} | {'sì' if column.nullable else 'no'} "
                 f"| {'sì' if column.pk else ''} | {cdesc} | {examples} |"
             )
-        if table.foreign_keys:
+        fks = table_foreign_keys(physical, annotations, table_name)
+        if fks:
             lines += ["", "Foreign keys:"]
-            for fk in table.foreign_keys:
+            for fk in fks:
                 lines.append(
                     f"- ({', '.join(fk.columns)}) → {fk.ref_table} ({', '.join(fk.ref_columns)})"
                 )

harness/tht/search/__init__.py

@@ -97,8 +97,12 @@ def combined_search(
     top: int,
     rrf_k: int,
     kinds: list[str] | None,
+    query_vec: list[float] | None = None,
 ) -> list[SearchResult]:
-    """Fonde LSH (valori di campo) e pgvector con Reciprocal Rank Fusion."""
+    """Fonde LSH (valori di campo) e pgvector con Reciprocal Rank Fusion.
+
+    `query_vec` permette di riusare un embedding gia' calcolato della stessa
+    keyword (es. `tht search pack`, che fa piu' ricerche sulla stessa domanda)."""
     rankings: dict[str, list[tuple[str, float]]] = {}
     lsh_values: dict[str, str] = {}
     if lsh_hits:
@@ -106,7 +110,9 @@ def combined_search(
         rankings["lsh"] = [(key, score) for key, score, _ in aggregated]
         lsh_values = {key: value for key, _, value in aggregated}
 
-    vector_hits = store.search(embedder.embed_query(keyword), top_n=top * 2, kinds=kinds)
+    if query_vec is None:
+        query_vec = embedder.embed_query(keyword)
+    vector_hits = store.search(query_vec, top_n=top * 2, kinds=kinds)
     rankings["vector"] = [(_vector_key(h), h.similarity) for h in vector_hits]
     by_key = {_vector_key(h): h for h in vector_hits}