File size: 21,037 Bytes
845e85a
feb1b1c
 
845e85a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
feb1b1c
845e85a
 
 
feb1b1c
845e85a
feb1b1c
845e85a
 
 
 
feb1b1c
845e85a
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604

from __future__ import annotations

import hashlib
import json
from collections.abc import Mapping
from enum import Enum
from types import MappingProxyType
from typing import final

from pydantic import BaseModel, ConfigDict, Field, model_validator

from redstack.domain.errors import ArtifactContractError
from redstack.domain.ids import FeatureIndex
from redstack.features.layout import (
    FEATURE_IDS,
    FEATURE_LAYOUT,
    GROUP_MEMBERS,
    GROUP_ORDER,
    LAYOUT_VERSION,
    SourceSlice,
)
from redstack.features.parsing import FeatureId, feature_id

_STRICT = ConfigDict(
    frozen=True, extra="forbid", str_strip_whitespace=True, validate_default=True
)


class UnknownFeatureId(KeyError):
    """Lookup of a feature id absent from the registry/layout."""


class FeatureTier(str, Enum):
    """Priority tier (Part 10): A mission-critical … D nice-to-have."""

    A = "A"
    B = "B"
    C = "C"
    D = "D"


class FeaturePolarity(str, Enum):
    """Whether a higher value is desirable, undesirable, or neutral."""

    POSITIVE = "positive"
    NEGATIVE = "negative"
    NEUTRAL = "neutral"


class FeatureDType(str, Enum):
    """CQV cell dtype. The value vector is uniformly float32 (flags are 0.0/1.0)."""

    FLOAT32 = "float32"


@final
class FeatureVersion(BaseModel):
    """Semantic version of one feature's extractor (Part 6 bump rules).

    Value-changing edit ⇒ minor bump; a layout/order change is a *global*
    ``layout_version`` major bump (owned by ``features.layout``).
    """

    model_config = _STRICT

    major: int = Field(ge=0)
    minor: int = Field(ge=0)
    patch: int = Field(ge=0)

    @property
    def semver(self) -> str:
        return f"{self.major}.{self.minor}.{self.patch}"

    def as_tuple(self) -> tuple[int, int, int]:
        return (self.major, self.minor, self.patch)


@final
class FeatureSchema(BaseModel):
    """dtype / range / nullability contract per feature; enforced at extraction.

    The CQV value vector carries no nulls — ``UNKNOWN`` is encoded as a neutral
    value plus low confidence — so ``nullable`` is ``False`` for every value
    feature.
    """

    model_config = _STRICT

    feature_id: FeatureId
    dtype: FeatureDType
    lower: float = Field(allow_inf_nan=False)
    upper: float = Field(allow_inf_nan=False)
    nullable: bool

    @model_validator(mode="after")
    def _bounds_ordered(self) -> FeatureSchema:
        if self.lower > self.upper:
            raise ValueError("FeatureSchema lower bound exceeds upper bound")
        return self


@final
class FeatureMetadata(BaseModel):
    """Human-facing metadata for reviewers + the O9 validation battery."""

    model_config = _STRICT

    doc: str = Field(min_length=1)
    expected_distribution: str = Field(min_length=1)
    owner: str = Field(min_length=1)
    tier: FeatureTier
    polarity: FeaturePolarity


@final
class FeatureDefinition(BaseModel):
    """The unit of the taxonomy: a fully-specified feature contract."""

    model_config = _STRICT

    feature_id: FeatureId
    group: str = Field(min_length=1)
    index: FeatureIndex = Field(ge=0)
    source_slice: SourceSlice
    schema_: FeatureSchema = Field(alias="schema")
    dependencies: tuple[FeatureId, ...]
    extractor_ref: str = Field(min_length=1)
    version: FeatureVersion
    tier: FeatureTier
    polarity: FeaturePolarity
    metadata: FeatureMetadata

    @model_validator(mode="after")
    def _consistent(self) -> FeatureDefinition:
        if self.schema_.feature_id != self.feature_id:
            raise ValueError("FeatureDefinition.schema feature_id mismatch")
        if self.metadata.tier is not self.tier:
            raise ValueError("FeatureDefinition tier/metadata.tier mismatch")
        if self.metadata.polarity is not self.polarity:
            raise ValueError("FeatureDefinition polarity/metadata.polarity mismatch")
        if not self.feature_id.startswith(f"{self.group}."):
            raise ValueError("FeatureDefinition group does not prefix feature_id")
        return self


@final
class FeatureManifest(BaseModel):
    """Per-run manifest: ``layout_version`` + ordered ids + group map + hash."""

    model_config = _STRICT

    layout_version: str = Field(min_length=1)
    feature_ids: tuple[FeatureId, ...] = Field(min_length=1)
    groups: Mapping[str, tuple[FeatureId, ...]]
    registry_hash: str = Field(min_length=1)

    def matches(self, layout_version: str) -> bool:
        """True iff this manifest was built for ``layout_version``."""
        return self.layout_version == layout_version

    def require_match(self, layout_version: str) -> None:
        """Raise ``ArtifactContractError`` unless versions agree (Scoring/O8)."""
        if not self.matches(layout_version):
            raise ArtifactContractError(
                f"layout_version mismatch: manifest={self.layout_version!r} "
                f"expected={layout_version!r}"
            )


# --------------------------------------------------------------------------- #
# Metadata priors: tier (Part 10), polarity (Part 1–5), extractor refs (§8).   #
# --------------------------------------------------------------------------- #

_EXTRACTOR_REF: Mapping[str, str] = MappingProxyType(
    {
        "id": "features.parsing",
        "geo": "features.geography.extract_geography",
        "reloc": "features.geography.extract_geography",
        "notice": "features.geography.extract_geography",
        "sal": "features.geography.extract_geography",
        "edu": "features.education.extract_education",
        "exp": "features.career",
        "sen": "features.career",
        "co": "features.career",
        "pvs": "features.career",
        "career": "features.career",
        "lead": "features.career",
        "startup": "features.career",
        "found": "features.career",
        "retr": "features.skills",
        "rank": "features.skills",
        "recsys": "features.skills",
        "ir": "features.skills",
        "nlp": "features.skills",
        "llm": "features.skills",
        "mle": "features.skills",
        "mlops": "features.skills",
        "eval": "features.skills",
        "oss": "features.skills",
        "cons": "features.skills",
        "avail": "features.signals",
        "eng": "features.signals",
        "resp": "features.signals",
        "bhv": "features.signals",
        "risk": "features.signals",
        "hp": "features.honeypot",
        "jd": "features.latents",
    }
)

# Features whose higher value is undesirable.
_NEGATIVE_IDS: frozenset[str] = frozenset(
    {
        "geo.outside_india_no_sponsor",
        "exp.derived_vs_stated_gap",
        "sen.title_vs_scope_gap",
        "pvs.consulting_density",
        "lead.management_only",
        "notice.over_30",
        "sal.is_inverted",
        "career.title_inflation",
        "career.consulting_density",
        "career.research_only",
        "career.management_only",
        "risk.uncertainty",
        "risk.contradiction",
        "bhv.behavioral_risk",
        "jd.keyword_only",
        "jd.consulting_only",
        "jd.title_chaser",
        "jd.pure_researcher",
        "jd.framework_enthusiast",
        "jd.inactive",
    }
    | {f"hp.{name}" for name in (
        "timeline_impossible", "skill_time_contradiction", "employment_overlap",
        "title_seniority_anomaly", "education_career_anomaly", "salary_anomaly",
        "experience_inflation", "keyword_stuffing", "behavioral_inconsistency",
        "signal_impossibility", "identity_anomaly", "composite",
    )}
)

# Features that are contextual / not directionally scored.
_NEUTRAL_IDS: frozenset[str] = frozenset(
    {"id.is_valid_id", "reloc.needed"}
    | {f"{group}.claimed" for group in (
        "retr", "rank", "recsys", "ir", "nlp", "llm", "mle", "mlops", "eval",
    )}
)

# Default tier by group; per-id overrides applied on top (Part 10).
_GROUP_TIER: Mapping[str, FeatureTier] = MappingProxyType(
    {
        "id": FeatureTier.C, "geo": FeatureTier.C, "exp": FeatureTier.C,
        "sen": FeatureTier.C, "edu": FeatureTier.C, "co": FeatureTier.C,
        "pvs": FeatureTier.A, "oss": FeatureTier.D, "lead": FeatureTier.D,
        "startup": FeatureTier.D, "found": FeatureTier.D, "avail": FeatureTier.C,
        "eng": FeatureTier.C, "resp": FeatureTier.C, "sal": FeatureTier.C,
        "reloc": FeatureTier.C, "notice": FeatureTier.C, "bhv": FeatureTier.C,
        "career": FeatureTier.C, "cons": FeatureTier.C, "risk": FeatureTier.C,
        "hp": FeatureTier.A, "jd": FeatureTier.A,
        "retr": FeatureTier.B, "rank": FeatureTier.B, "recsys": FeatureTier.B,
        "ir": FeatureTier.B, "nlp": FeatureTier.B, "llm": FeatureTier.B,
        "mle": FeatureTier.B, "mlops": FeatureTier.B, "eval": FeatureTier.B,
    }
)

_TIER_OVERRIDE: Mapping[str, FeatureTier] = MappingProxyType(
    {
        # pvs
        "pvs.product_density": FeatureTier.A,
        "pvs.product_recent": FeatureTier.A,
        "pvs.consulting_density": FeatureTier.B,
        # experience
        "exp.in_band": FeatureTier.B,
        # career intelligence (Part 10 Tier B + the JD-critical density)
        "career.product_company_density": FeatureTier.A,
        "career.production_exposure": FeatureTier.B,
        "career.technical_depth": FeatureTier.B,
        "career.hands_on_engineering": FeatureTier.B,
        "career.title_inflation": FeatureTier.B,
        "career.consulting_density": FeatureTier.B,
        # behavioral multiplier inputs
        "bhv.availability": FeatureTier.B,
        "bhv.hiring_probability_proxy": FeatureTier.B,
        "bhv.market_momentum": FeatureTier.D,
        "bhv.engagement_velocity": FeatureTier.D,
        # honeypot soft / nuance
        "hp.salary_anomaly": FeatureTier.C,
        "hp.behavioral_inconsistency": FeatureTier.B,
        # jd negatives (Part 10 Tier B)
        "jd.consulting_only": FeatureTier.B,
        "jd.pure_researcher": FeatureTier.B,
        "jd.framework_enthusiast": FeatureTier.B,
        "jd.title_chaser": FeatureTier.B,
        "jd.inactive": FeatureTier.B,
        "jd.shipping_mentality": FeatureTier.B,
    }
)

# Competency suffix → tier (anti-stuffer trust/competency are mission-critical;
# raw claimed is near-zero weight).
_COMPETENCY_SUFFIX_TIER: Mapping[str, FeatureTier] = MappingProxyType(
    {
        "claimed": FeatureTier.D,
        "trust": FeatureTier.A,
        "in_career": FeatureTier.B,
        "semantic": FeatureTier.B,
        "competency": FeatureTier.A,
    }
)

_COMPETENCY_GROUPS: frozenset[str] = frozenset(
    {"retr", "rank", "recsys", "ir", "nlp", "llm", "mle", "mlops", "eval"}
)

_GROUP_DISTRIBUTION: Mapping[str, str] = MappingProxyType(
    {
        "career": "right-skewed-low (synthetic-noisy pool)",
        "hp": "spike-at-0, thin impossible tail",
        "pvs": "bimodal (product vs services)",
        "jd": "regressed-to-prior under low confidence",
    }
)


def _tier_for(group: str, name: str, fid: str) -> FeatureTier:
    if fid in _TIER_OVERRIDE:
        return _TIER_OVERRIDE[fid]
    if group in _COMPETENCY_GROUPS:
        return _COMPETENCY_SUFFIX_TIER[name]
    return _GROUP_TIER[group]


def _polarity_for(fid: str) -> FeaturePolarity:
    if fid in _NEGATIVE_IDS:
        return FeaturePolarity.NEGATIVE
    if fid in _NEUTRAL_IDS:
        return FeaturePolarity.NEUTRAL
    return FeaturePolarity.POSITIVE


def _dependencies_for(fid: str) -> tuple[FeatureId, ...]:
    return _DEPENDENCIES.get(fid, ())


def _competency(group: str, suffix: str) -> FeatureId:
    return feature_id(group, suffix)


# Latent / composite dependency wiring (Part 2 constituents; hp.composite over
# its detectors). Primitives carry no dependencies.
_DEPENDENCIES: Mapping[str, tuple[FeatureId, ...]] = MappingProxyType(
    {
        "jd.retrieval_ranking": (
            _competency("retr", "competency"),
            _competency("rank", "competency"),
            _competency("ir", "competency"),
        ),
        "jd.production_ml": (
            _competency("mle", "competency"),
            _competency("mlops", "competency"),
            feature_id("pvs", "product_density"),
        ),
        "jd.product_company": (
            feature_id("pvs", "product_density"),
            feature_id("pvs", "product_recent"),
        ),
        "jd.shipping_mentality": (
            feature_id("startup", "shipping_signal"),
            feature_id("found", "ownership"),
        ),
        "jd.eval_framework": (_competency("eval", "competency"),),
        "jd.hybrid_retrieval": (
            _competency("ir", "competency"),
            _competency("ir", "trust"),
        ),
        "jd.keyword_only": tuple(
            _competency(group, suffix)
            for group in ("retr", "rank", "recsys", "ir", "nlp", "llm", "mle", "mlops", "eval")
            for suffix in ("claimed", "trust", "in_career", "semantic")
        ),
        "jd.consulting_only": (
            feature_id("pvs", "consulting_density"),
            feature_id("career", "consulting_density"),
        ),
        "jd.title_chaser": (
            feature_id("career", "title_inflation"),
            feature_id("career", "stability"),
        ),
        "jd.pure_researcher": (
            feature_id("career", "research_only"),
            feature_id("career", "production_exposure"),
        ),
        "jd.framework_enthusiast": (
            _competency("llm", "competency"),
            feature_id("career", "hands_on_engineering"),
        ),
        "jd.inactive": (
            feature_id("avail", "available"),
            feature_id("resp", "reliable"),
            feature_id("bhv", "availability"),
        ),
        "hp.composite": tuple(
            feature_id("hp", name)
            for name in (
                "timeline_impossible", "skill_time_contradiction", "employment_overlap",
                "title_seniority_anomaly", "education_career_anomaly", "salary_anomaly",
                "experience_inflation", "keyword_stuffing", "behavioral_inconsistency",
                "signal_impossibility", "identity_anomaly",
            )
        ),
    }
)


def _build_definition(entry_name: str) -> FeatureDefinition:
    entry = next(e for e in FEATURE_LAYOUT.entries if e.name == entry_name)
    fid = FeatureId(entry.name)
    group, name = entry.name.split(".", 1)
    tier = _tier_for(group, name, entry.name)
    polarity = _polarity_for(entry.name)
    schema = FeatureSchema(
        feature_id=fid,
        dtype=FeatureDType.FLOAT32,
        lower=entry.lower,
        upper=entry.upper,
        nullable=False,
    )
    metadata = FeatureMetadata(
        doc=f"{entry.name}: {group}-group feature (source slice {entry.source_slice}).",
        expected_distribution=_GROUP_DISTRIBUTION.get(group, "unit[0,1]"),
        owner=_EXTRACTOR_REF[group],
        tier=tier,
        polarity=polarity,
    )
    return FeatureDefinition(
        feature_id=fid,
        group=group,
        index=entry.index,
        source_slice=SourceSlice(entry.source_slice),
        schema=schema,
        dependencies=_dependencies_for(entry.name),
        extractor_ref=_EXTRACTOR_REF[group],
        version=FeatureVersion(major=1, minor=1, patch=0),
        tier=tier,
        polarity=polarity,
        metadata=metadata,
    )


def _canonical_blob(definitions: tuple[FeatureDefinition, ...]) -> str:
    """A deterministic, order-stable serialization for the registry hash."""
    rows = [
        {
            "i": int(d.index),
            "id": str(d.feature_id),
            "group": d.group,
            "slice": d.source_slice.value,
            "lo": d.schema_.lower,
            "hi": d.schema_.upper,
            "dtype": d.schema_.dtype.value,
            "nullable": d.schema_.nullable,
            "tier": d.tier.value,
            "polarity": d.polarity.value,
            "ver": d.version.semver,
            "deps": [str(dep) for dep in d.dependencies],
        }
        for d in definitions
    ]
    return json.dumps(
        {"layout_version": LAYOUT_VERSION, "features": rows},
        sort_keys=True,
        separators=(",", ":"),
        ensure_ascii=True,
    )


@final
class FeatureRegistry(BaseModel):
    """The frozen set of all ``FeatureDefinition``s; id↔index source of truth."""

    model_config = ConfigDict(frozen=True, extra="forbid", validate_default=True)

    definitions: tuple[FeatureDefinition, ...] = Field(min_length=1)
    layout_version: str = Field(min_length=1)

    @model_validator(mode="after")
    def _matches_layout(self) -> FeatureRegistry:
        if self.layout_version != LAYOUT_VERSION:
            raise ArtifactContractError(
                f"registry layout_version {self.layout_version!r} != "
                f"layout {LAYOUT_VERSION!r}"
            )
        if len(self.definitions) != len(FEATURE_IDS):
            raise ArtifactContractError("registry size != layout dim")
        seen: set[str] = set()
        for position, definition in enumerate(self.definitions):
            if int(definition.index) != position:
                raise ArtifactContractError("registry definitions out of index order")
            if str(definition.feature_id) != FEATURE_IDS[position]:
                raise ArtifactContractError(
                    f"registry id {definition.feature_id!r} != layout "
                    f"{FEATURE_IDS[position]!r} at index {position}"
                )
            if str(definition.feature_id) in seen:
                raise ArtifactContractError("duplicate feature id in registry")
            seen.add(str(definition.feature_id))
        # Every declared dependency must itself be a known feature id.
        for definition in self.definitions:
            for dep in definition.dependencies:
                if str(dep) not in seen:
                    raise ArtifactContractError(
                        f"dependency {dep!r} of {definition.feature_id!r} is unknown"
                    )
        return self

    @property
    def dim(self) -> int:
        return len(self.definitions)

    def get(self, feature_id_: FeatureId) -> FeatureDefinition:
        """Return the definition for ``feature_id_``; raise on unknown id."""
        index = self._index_of.get(str(feature_id_))
        if index is None:
            raise UnknownFeatureId(str(feature_id_))
        return self.definitions[index]

    def has(self, feature_id_: FeatureId) -> bool:
        return str(feature_id_) in self._index_of

    def index(self, feature_id_: FeatureId) -> FeatureIndex:
        """Return the CQV index of ``feature_id_``; raise on unknown id."""
        return self.get(feature_id_).index

    def feature_id_at(self, index: FeatureIndex) -> FeatureId:
        """Return the feature id occupying ``index``."""
        position = int(index)
        if position < 0 or position >= len(self.definitions):
            raise UnknownFeatureId(f"index {position} out of range")
        return self.definitions[position].feature_id

    def by_group(self, group: str) -> tuple[FeatureDefinition, ...]:
        """All definitions in ``group`` in layout order."""
        return tuple(d for d in self.definitions if d.group == group)

    @property
    def groups(self) -> tuple[str, ...]:
        return GROUP_ORDER

    @property
    def registry_hash(self) -> str:
        """Stable sha256 of the canonical definition serialization."""
        blob = _canonical_blob(self.definitions)
        return hashlib.sha256(blob.encode("utf-8")).hexdigest()

    def manifest(self) -> FeatureManifest:
        """Build the per-run ``FeatureManifest`` (ids, group map, hash)."""
        groups: Mapping[str, tuple[FeatureId, ...]] = MappingProxyType(
            {
                group: tuple(FeatureId(fid) for fid in GROUP_MEMBERS[group])
                for group in GROUP_ORDER
            }
        )
        return FeatureManifest(
            layout_version=self.layout_version,
            feature_ids=tuple(FeatureId(fid) for fid in FEATURE_IDS),
            groups=groups,
            registry_hash=self.registry_hash,
        )

    @property
    def _index_of(self) -> Mapping[str, int]:
        # Recomputed cheaply (frozen model; small, no cached mutable state).
        return MappingProxyType(
            {str(d.feature_id): int(d.index) for d in self.definitions}
        )


def _build_registry() -> FeatureRegistry:
    definitions = tuple(_build_definition(name) for name in FEATURE_IDS)
    return FeatureRegistry(definitions=definitions, layout_version=LAYOUT_VERSION)


# The single, frozen registry instance shared by reference (memory §Q).
FEATURE_REGISTRY: FeatureRegistry = _build_registry()
FEATURE_MANIFEST: FeatureManifest = FEATURE_REGISTRY.manifest()


__all__: tuple[str, ...] = (
    "FEATURE_MANIFEST",
    "FEATURE_REGISTRY",
    "FeatureDType",
    "FeatureDefinition",
    "FeatureManifest",
    "FeatureMetadata",
    "FeaturePolarity",
    "FeatureRegistry",
    "FeatureSchema",
    "FeatureTier",
    "FeatureVersion",
    "UnknownFeatureId",
)