feat: Week 2 — brain/emotion package (7 modules) by hanamorix · Pull Request #2 · hanamorix/companion-emergence

hanamorix · 2026-04-22T14:01:50Z

Summary

Ships the full brain/emotion/ package per spec Section 5 (the emotional core — P1 organising principle)
7 sub-modules: vocabulary, state, decay, arousal, blend, influence, expression
26-baseline emotion taxonomy + persona extension registry
75 new tests; total suite now 113 across macOS + Windows + Linux

Test plan

pytest — 113 tests pass locally
ruff check + format — clean
Manual smoke: all 7 sub-modules compose correctly in a single flow
CI matrix green across all 3 OSes (verifies after push)

8 tasks, 54 bite-sized checkbox steps, TDD throughout. Builds the brain/emotion/ package per spec Section 5 (P1 — the emotional core as organising principle). Scope: 7 sub-modules in the package — - vocabulary (26-baseline Emotion taxonomy + persona extensions) - state (EmotionalState + ResidueEntry with clamping/dominant/residue) - decay (per-emotion half-life, spec-pinned grief=60d joy=3d identity=None) - arousal (7-tier spectrum from state + body temperature) - blend (emergent co-occurrence detection, pairs + triples, ≥5 threshold) - influence (state → structured biasing hints for providers, Week 5) - expression (state → 24 facial + 8 arm/hand params, NellFace Week 6) Integration smoke in Task 8 exercises all seven modules composed. Expected test count post-Week-2: 107 (Week 1's 38 + 69 new). Out of scope (deferred): bridge provider integration (Week 5), voice module (Week 4), engines (Week 4), memory + migrator (Week 3). Execution via superpowers:subagent-driven-development. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…istry Ships the baseline 26-emotion taxonomy (11 core + 10 complex + 5 Nell-specific) with typed Emotion dataclass (name, description, category, half-life, clamp). Half-lives per spec Section 10.1: grief=60d, joy=3d, anchor_pull/love/belonging/ body_grief/freedom_ache=None (identity-level); others seeded for later tuning. register()/_unregister() support per-persona extensions. 13 tests green. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The Task 1 test file has 13 test functions; I had "12" in several places in the plan and its cascading "expected: N passed" counts downstream. Implementer caught this in DONE_WITH_CONCERNS; amended their commit message to say "13 tests green" and corrected the plan throughout: - Task 1: 12 → 13 - Task 1 cumulative: 50 → 51 - Task 2-7 cumulative: each +1 (62, 70, 79, 88, 97, 108) - Task 8 green-light total: 107 → 108 - Task 8 PR body: 69 new tests → 70 new tests Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

From the code quality reviewer's findings: Important: - Emotion.category typed as Literal["core", "complex", "nell_specific", "persona_extension"] via EmotionCategory type alias. Edit-time typo catching via mypy/pyright; by_category() silent-empty-on-typo stops being a silent footgun. No runtime check added (YAGNI). - intensity_clamp: int → float (default 10.0). Prevents silent int/float coercion when Task 2 state.py compares float intensities against the clamp. No behaviour change for existing values (all ints coerce fine). - Emotion attribute docstring updated to include freedom_ache in the identity-level examples list (was previously missing; module docstring already listed it correctly). - Fix test_emotion_half_life_may_be_none to pass a valid category "nell_specific" instead of the stray "persona" string. No runtime impact but matches the new Literal annotation and avoids confusion. Minor: - Thread-safety comment on _REGISTRY noting extensions must register at startup before any concurrent reader. Prevents a future "why is my registry corrupt" debug session once the async bridge lands. Deliberately NOT applied: - _unregister → @contextmanager: premature at one caller, revisit Task 7. - ValueError → custom EmotionAlreadyRegistered exception: no caller needs to differentiate; ValueError is adequate. - Half-life seed re-tuning: values are spec-explicit "seed/tunable" by design; adjust based on lived experience, not speculative review. Pytest: 51/51 pass. Ruff + ruff format both clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

EmotionalState holds {emotion: intensity}, a dominant pointer, and a bounded residue queue for temporal carry-over. All writes validated against the vocabulary (clamped, non-negative, known emotions only). to_dict/from_dict round-trips cleanly. 11 tests green. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

From the code quality reviewer's findings: Important: - EmotionalState.__post_init__ added — calls _recompute_dominant() so direct dataclass construction (which later tasks will use for snapshots and intermediate states) self-heals a stale dominant pointer. Prevents silent (emotions vs dominant) inconsistency. - from_dict docstring now documents permissive-by-design behaviour: unknown emotion names preserved as-is to support schema migration and retired-vocabulary tolerance. Downstream consumers (decay in Task 3) handle unknown names via vocabulary.get() returning None — already covered by Task 3's test_decay_ignores_unknown_emotion_in_state. - ResidueEntry.from_dict now coerces tz-naive timestamps to UTC rather than silently accepting them. Matches the docstring's "UTC-aware" claim and prevents future tz-comparison landmines. Minor: - copy() gains a MAINTENANCE note — future EmotionalState fields must be copied explicitly; intentionally manual to keep shallow/deep decisions visible per field. - add_residue() docstring notes vocabulary validation is intentionally skipped on residue entries (historical record may contain retired emotion names). - to_dict() comment notes `dominant` is serialised for human readability of the JSON; from_dict ignores it and recomputes for self-consistency. Deliberately NOT applied: - deque(maxlen=N) for residue instead of list slicing: O(N) eviction at residue_max=16 is negligible; dataclass-field simplicity wins. - Full validation pass in from_dict: the permissive design is deliberate (now documented) and downstream consumers handle it. Pytest: 62/62 pass. Ruff + ruff format both clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

apply_decay(state, elapsed_seconds) applies exponential decay to every known emotion in state using its vocabulary half-life. Identity-level emotions (half_life=None) are untouched. Below the noise floor (0.01) emotions are removed entirely. 8 tests green. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

From the code quality reviewer's findings: Important: - Added test_partial_decay_matches_exponential_formula — guards against a regression to linear decay. The two exactly-at-half-life tests (grief/joy) would pass a linear implementation too; this pins the curve shape at a non-boundary point. Minor: - apply_decay docstring now documents the silent no-op behaviour for non-positive elapsed_seconds (keeping the current semantic, just making the contract explicit for callers). - test_decay_ignores_unknown_emotion_in_state now asserts state.dominant == "unknown_emotion" — makes the intentional side effect (permissive-contract data becomes dominant) explicit rather than surprising. Deliberately NOT applied: - Promote _recompute_dominant() to public API. Reviewer's forward-looking concern was "this spreads to Tasks 4-6". Checked the plan for 4-7: all are read-only consumers of state; none call _recompute_dominant. decay remains the only outside caller; one case doesn't warrant API surface expansion. Reconsider if Tasks 4-7 grow write surface later. - Strict-< on noise floor: floating-point arithmetic essentially never produces exactly 0.01, so behaviour is deterministic in practice. Pytest: 71/71 pass (70 + 1 new partial-decay test). Ruff + format clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Tiers: dormant → casual → warmed → reaching → charged → held → edge. compute_tier() maps emotional state + body temperature to a discrete tier. Suppressors (grief, shame, fear) reduce arousal; body-temp warmth amplifies. Week 2 scope: forward direction only (state→tier); reverse coupling in Week 4. 9 tests green. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Plan shipped with seed weights desire=0.8, tenderness=0.4 which produced raw=10.1 for the desire=8+tenderness=7+body_temp=3 test, landing at HELD (5) outside the test's expected (REACHING, CHARGED) range. Implementer caught this in DONE_WITH_CONCERNS; adjusted to desire=0.7, tenderness=0.2 which math out correctly (raw=7.9 → CHARGED). Plan updated to match. All 9 arousal tests pass with the calibrated weights; full 80-test suite is green; lint clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

From the code quality reviewer's findings: Important: - Love weight 0.2 → 0.15 to preserve the module docstring's semantic contract: "Love alone doesn't progress past warmed." At weight=0.2 and max intensity 10, raw=2.0 exactly — and the threshold chain uses `raw < 2.0` which falls through to REACHING. At weight=0.15, raw=1.5 stays safely inside WARMED. Verified all other tests still pass. - Post-suppression zero now returns TIER_DORMANT rather than leaking into CASUAL via the <0.5 threshold. "Desire crushed by grief" semantically maps to "no signal" (DORMANT = "no arousal signal at all" per the tier docstring), not "everyday warmth". Minor: - _AROUSAL_EMOTIONS and _SUPPRESSORS wrapped in types.MappingProxyType. Read-only dict view; prevents accidental runtime mutation via `arousal._AROUSAL_EMOTIONS["desire"] = 9000`. Stdlib, zero runtime cost, iteration/get() identical to dict. - test_body_temperature_ignored_when_no_arousal_source: docstring and assertion updated from "casual range" to the actually-correct DORMANT (body temp alone cannot create arousal — the early return fires). Test now asserts `== TIER_DORMANT` rather than `<= WARMED`. Boundary-value tests added: - test_max_love_alone_still_in_warmed (guards the weight calibration against future love-weight drift) - test_suppression_to_zero_returns_dormant (guards the new post- suppression DORMANT semantic against future refactor) Plan updated to match (love weight 0.2 → 0.15 + explanatory comment). Deliberately NOT applied: - IntEnum for tier constants — plain int is idiomatic and tests rely on direct comparison; IntEnum adds weight for marginal gain. - body_temperature: int → float — int is consistent with the OG's body_state.json format; widen when Week 4 body-state coupling lands. - Boundary tests for every threshold — the two added guards cover the cases that would bite downstream consumers; more can come in Week 4. Pytest: 82/82 pass (80 + 2 boundary tests). Ruff + format both clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

BlendDetector observes EmotionalState instances over time, counting which emotion combinations repeatedly co-occur at high intensity. Pair + triple combinations tracked; threshold ≥5 intensity each, ≥5 co-occurrences to register. name_blend() labels a detected pattern. Round-trips through to_dict/from_dict. 9 tests green. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

From the code quality reviewer's findings: Important: - from_dict now sorts component tuples on load. observe() and name_blend() always operate on sorted keys; from_dict previously stored whatever order the JSON had, which would cause spurious KeyErrors on subsequent name_blend() calls if a hand-edited config had unsorted components. One tuple(sorted(...)) per loop; no behaviour change on well-formed to_dict output but defence in depth for external data. - test_name_unknown_blend_raises replaced try/except/else pattern with pytest.raises(KeyError, match="has not been detected"). The old form's first check ("not detected" in message) was dead code — the actual message is "has not been detected yet", so substring lookup failed; the test passed only because the second branch ("love" in key repr) was true. Now checks the real phrase so a future error-message drift would be caught. Minor: - _observations and _names fields use init=False so they don't appear in the public constructor signature. from_dict assigns post-construction (unchanged from before); external callers can no longer inject arbitrary counts via BlendDetector(_observations={...}). - _names type annotation widened to dict[tuple[str, ...], str | None] (never actually holds None in practice — from_dict now skips None names explicitly — but the type matches the JSON null possibility). - Size-3 cap comment expanded to explain the rationale: meaningful emergent patterns ("building_love", "creative_feral") are typically pair or triple shaped; higher-order blends dilute and rarely register. Rule tunable if real data accrues a different pattern. Deliberately NOT applied: - __init__.py exports for BlendDetector: plan explicitly says "consumers import from brain.emotion directly" — and the __all__ question for the package root belongs to Task 8 (close-out). No change now. - Explicit test for (triple + pair co-detection = 4 blends): this is known intentional behaviour; a regression test would be nice-to-have, not required before Tasks 6/7 which don't consume blend directly. Pytest: 91/91 pass. Ruff + format both clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

calculate_influence(state, arousal_tier, energy) returns an InfluenceHints dataclass with dominant emotion, arousal tier, tone bias, voice register, and suggested length multiplier. Provider-agnostic — bridge providers in Week 5 decide how to render each field. 9 tests green. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

From the code quality reviewer's findings: Important: - Removed "warm" from voice_register docstring — never produced by the code. Stale label from an earlier design pass; a Week 5 consumer pattern-matching on "warm" would get a dead branch. Actual produced values are default/soft/intimate/terse. - _TONE_RULES list → tuple of tuples. Same mutability footgun pattern flagged in Tasks 1 and 4; converting to immutable tuple prevents accidental runtime mutation of rule matching. - Documented the length multiplier chain behaviour explicitly. TIER_HELD clamps into a deliberate-pace band [0.8, 1.2]; TIER_EDGE hard-overrides to 0.8. The chain's non-obvious effect (crisp at HELD = 0.8, generative at HELD = 1.2) is now documented via inline comments rather than left as implicit emergence. Minor: - test_hints_to_dict_round_trips now asserts suggested_length_multiplier round-trips. Previous test was missing that field's check; a future accidental removal of the key would have been undetected. - Added test_held_tier_clamps_length_into_deliberate_band (covers the generative-down and crisp-up clamping behaviour at TIER_HELD). - Added test_edge_tier_hard_overrides_length_to_terse (covers the absolute 0.8 override regardless of tone_bias). Deliberately NOT applied: - state.emotions.get(dominant, 0.0) → state.emotions[dominant]: reviewer flagged the fallback as unreachable, but defensive .get is low-cost and signals "we're paranoid about state invariants" — keep. - Voice register precedence test (grief + TIER_CHARGED scenario): reviewer flagged as design decision to document. The current "arousal wins over grief" behaviour is intentional; a test would just assert the current semantic without adding value. Defer to lived experience. Pytest: 102/102 pass (100 + 2 new HELD/EDGE boundary tests). Ruff + format both clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

compute_expression(state, arousal_tier, energy) returns an ExpressionVector with 24 facial params + 8 arm/hand params, ready for NellFace renderer consumption in Week 6. Art-agnostic — outputs numbers; expression_map.json decides how to render against them. Parameter ramps per-emotion (joy, grief, anger, fear, tenderness, desire) with arousal-tier-gated body-tension escalation. All facial params clamped to [0,1]; hand_pose is an enum string. 11 tests green. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

From the code quality reviewer's findings: Important: - Added _HAND_POSES constant — 9-pose enum tuple (resting, reaching, holding, gesturing, clasped, writing, guarded, open, fist). Week 2's compute_expression produces 4 of these; the others are defined for NellFace animation authoring in Week 6. ExpressionVector docstring updated to reference _HAND_POSES instead of "a small enum". - test_all_params_in_zero_to_one_range now asserts float-range [0.0, 1.0] on arm_hand values (skipping hand_pose which is str). Previous test only verified type, not range — coverage gap. Minor (comments only): - Inline comment on the grief mouth coefficient (0.5 vs joy's 0.4): intentional asymmetry — mixed states land below neutral, matching human pattern where sadness dominates mouth expression. - Inline comment on body_heat = max(desire, arousal_emotion): arousal decays in hours (half-life 0.5d), desire decays in days. max() keeps the transient arousal signal from being washed out. - Docstring note on compute_expression: Week 2 only drives arousal- tier-gated arm params; arm_openness/wrist_angle/finger_spread/ reach_retract stay at baseline 0.3. NellFace Week 6 handles those via pose-driven animation. Deliberately NOT applied: - brow_furrow saturation at grief+anger extremes: known tuning issue for Week 6 ramp-weight refinement. - _clamp(0.7) vacuous-clamp on literal constant: cosmetic only. Pytest: 113/113 pass. Ruff + format both clean. This closes out all 7 emotion sub-modules. Task 8 (week close-out + tag week-2-complete) is next. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

… msg - HeartbeatEngine.reflex_arcs_path and reflex_log_path now default to None instead of bare cwd-relative Path("reflex_arcs.json"). When either is None, _try_fire_reflex short-circuits with an empty result — no more silent writes to cwd from tests that don't explicitly configure reflex (items #2 and #4 from spec §15) - CLI heartbeat handler now distinguishes first-tick + --dry-run ('Would initialize on first real tick — work deferred.') from first-tick + live ('Heartbeat initialized — work deferred until next tick.') Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Spec §9.1 expanded from a one-line deferral into a concrete plan the next engineer can implement directly when soul module lands. Covers: - file classification (atomic-rewrite identity, same tier as emotion_vocabulary.json) - reconstruct_soul_from_memories(store) following F37's self-claims- from-experience pattern - schema validator shape - acceptance criteria for the soul-module PR Inline comments in walker.py (_DEFAULTS) and alarm.py (_IDENTITY_FILES) point at spec §9.1 so the plan is visible during code-reading too. Closes followup #2 from brain-health-module-design.md §9.

… + growth anomaly collector (#20) * feat(health): wire reconstruct_vocabulary_from_memories into vocabulary heal flow (F1) When emotion_vocabulary.json corrupts and no .bak is recoverable, the heal flow used to reset to empty `{"version":1,"emotions":[]}` — losing all persona-extension emotions the brain has been operating with. Now: if the loader has store access (caller passed `store=...`), the reset_to_default path is replaced with reconstruct_from_memories. The brain re-learns its own vocabulary from how it has been using emotions. Framework baseline (21 entries) + persona-extension entries detected in memories.db (with `(reconstructed from memory)` placeholder description and conservative 1.0-day decay). The anomaly's action field reflects the actual outcome: when reconstruction fires, action becomes `reconstructed_from_memories` (not `reset_to_default`). The forensic quarantine of the original corrupt file is preserved. Falls back to bare reset when no store is provided (some callers don't have one — that's fine, they'll get the empty default). Closes followup #1 from brain-health-module-design.md §9. * docs(health): concrete soul module health plan (F2) Spec §9.1 expanded from a one-line deferral into a concrete plan the next engineer can implement directly when soul module lands. Covers: - file classification (atomic-rewrite identity, same tier as emotion_vocabulary.json) - reconstruct_soul_from_memories(store) following F37's self-claims- from-experience pattern - schema validator shape - acceptance criteria for the soul-module PR Inline comments in walker.py (_DEFAULTS) and alarm.py (_IDENTITY_FILES) point at spec §9.1 so the plan is visible during code-reading too. Closes followup #2 from brain-health-module-design.md §9. * feat(health): thread anomaly collector through run_growth_tick (F3) When run_growth_tick reads a corrupt emotion_vocabulary.json via _read_current_vocabulary_names, the anomaly produced by the heal flow is now appended to an optional caller-provided collector instead of being silently dropped after a local warning. Wiring: - _read_current_vocabulary_names returns (set[str], BrainAnomaly | None) - run_growth_tick accepts anomalies_collector: list[BrainAnomaly] | None - HeartbeatEngine._try_run_growth forwards tick_anomalies as the collector when calling run_growth_tick After this lands, vocabulary corruption discovered inside the weekly growth tick surfaces in the heartbeat audit log + HeartbeatResult.anomalies + compact CLI 🩹/banner exactly like config/state corruption discovered at the top of the tick. No more silent loss. Calling run_growth_tick standalone (e.g., from tests, or in the future from a scheduled job runner) without a collector still works — the parameter is opt-in. Closes followup #3 from brain-health-module-design.md §9. --------- Co-authored-by: Hana <hana@nanoclaw.local>

SP-7 wraps SP-6's chat engine in a per-persona FastAPI daemon on localhost (dynamic port). Folds the conversation supervisor as a non-daemon thread for close_stale_sessions ticks, broadcasts brain events over WebSocket for Tauri/CLI subscribers, and ships dirty- shutdown recovery via a shutdown_clean flag in bridge.json. Resolves master-ref §8 open question #1 (transport: HTTP+WS, mirroring OG). Defers #2-#4 to SP-8, scopes #5-#6 to SP-6, marks #7-#8 unrelated. Folds in three audit must-fixes: shutdown_clean recovery, EventBus thread-safety with drop-on-overflow, pinned close_stale_sessions params. Six implementation chunks with smoke-test gates at each boundary; ~26 tests targeted (10 unit + 16 integration). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The shoji-styled left column from the mockups (mock-ups/app-interface/ nell_face_example_1..5.png), all reading the same /persona/state poll already wired in Phase 2. ## What landed Brain side: /persona/state gains a connection block: { provider, model, last_heartbeat_at } provider read from PersonaConfig.provider, model resolved via a v1 default-per-provider table (claude-cli → sonnet, ollama → the qwen2.5-abliterated default, fake → fake), last_heartbeat_at read from heartbeat_state.json. Fail-soft: missing files → null. App side (app/src/components/): - ui.tsx shared primitives — Bar (with progress fill + label), SectionLabel, Divider, Toggle, PanelShell. - panels/ InnerWeatherPanel.tsx top-N emotions sorted desc + body energy/ temp summary (mockup #1). BodyPanel.tsx full body block — energy/temp/exhaustion + body_emotions filtered >0.4 + session/ contact metadata (mockup #2). InteriorPanel.tsx dream/research/heartbeat/reflex paragraphs (mockup #3). Absent sections render nothing — silence is meaningful. SoulPanel.tsx one crystallization quote with love_type tag, resonance, date, why_it_matters. ConnectionPanel.tsx bridge mode + provider/model/heartbeat, integrations toggles (Obsidian/IPC stubbed disabled), window settings (always-on-top + reduced-motion live-toggleable). - LeftPanel.tsx container with icon-column tab switcher, renders the active panel; matches the mockup's small icon stack near the avatar. - styles.css user-controlled reduced-motion via data-reduced-motion="true" on <html>. - App.tsx wires LeftPanel + always-on-top + reduced-motion state. ## Phase 4-5 still ahead - Phase 4: install wizard as a separate Tauri window, ported from mock-ups/wizard-interface/Nell Wizard.html. Bridge auto-spawn from the app side (so users don't have to run nell supervisor manually). Persona selection persistence. - Phase 5: refine emotion-vector → expression mapping; per-persona face catalogue overrides; expression variant rotation on idle. Tests: 1386 -> 1388 (+2 — connection block populated + missing-file safe). All builds green: tsc, vite, cargo.

…ll launcher Two bugs surfaced during Hana's wizard validation against the Phase 7 bundled .app: 1. install_voice_template tried to read docs/voice-drafts/nell-voice.md from a path resolved relative to brain/setup.py — which is correct in source-installs but FAILS in any wheel install (the wheel ships only the brain/ package, not docs/). The nell-example voice template threw FileNotFoundError on every install_only path including the Phase 7 .app. 2. The pip-generated nell entry point at python-runtime/bin/nell had its python interpreter shebang baked to an absolute path from the build machine. The .app shipped a script that tried to exec /Users/<builder>/.../python-runtime/bin/python3 — a path that exists on the build machine and fails everywhere else. Hana's traceback shows the source-tree's brain/cli.py getting imported instead of the bundled one because the shebang resolved to the build host's python. Fix #1 — package the template: * docs/voice-drafts/nell-voice.md moved to brain/voice_templates/nell-voice.md * brain/voice_templates/__init__.py added so it's a real package (hatchling auto-includes everything under brain/) * install_voice_template reads via importlib.resources files('brain.voice_templates').joinpath('nell-voice.md'); no repo_root parameter, no path resolution * verified: wheel ships brain/voice_templates/nell-voice.md; nell init --voice-template nell-example writes the file correctly from a tmp NELLBRAIN_HOME Fix #2 — relocatable nell launcher: * build_python_runtime.sh post-install step replaces bin/nell with a 6-line /bin/sh wrapper that resolves $SCRIPT_DIR via cd "$(dirname "$0")" then execs "$SCRIPT_DIR/python3" with a -c 'from brain.cli import main; sys.exit(main())' payload. Whatever directory the python-runtime tree gets copied to, bin/nell finds its own python next door * Windows: pip's Scripts/nell.exe is already a relative-path launcher binary, so the script keeps it untouched * verified: bundled bin/nell runs --version from /tmp, init --voice-template nell-example writes voice.md Test updates: * test_install_voice_template_nell_example_copies_packaged_file replaces the old fake-repo-root test; asserts the canonical nell voice opens with '## 1. Who you are' and is > 1000 bytes * test_install_voice_template_nell_example_missing_file_raises removed (the importlib.resources path can't fail this way once the file ships in the wheel; ValueError on unknown template still tested) 1470 → 1469 (one removed test). Ruff clean. macOS arm64 .app re-bundled at ~/wizard-validation/ with codesign verify pass.

feat: Week 2 — brain/emotion package (7 modules)