feat: profiles page + container edit drawer + resolved_command (#658)

[thinmintdev]

Summary

Implements issue #658 — Profiles page, container-slot edit drawer, and backend resolved_command.

Backend (src/hal0/providers/container.py, src/hal0/api/routes/slots.py)

• resolved_command_for_slot(): new pure helper that builds the real llama-server argv for a container slot (image tag + --host, --port, --model, profile flags). Omits the podman preamble (requires root/GIDs, not useful for debugging).
• Model extraction: correctly reads from cfg["model"]["default"] (nested TOML table) — was erroneously str(cfg["model"]) which emitted a dict repr like {'default': 'llama-3b'}.
• _container_state_enrichment(): emits runtime, profile, image, resolved_command on every container slot list entry. Applied on both the list (GET /api/slots) and detail (GET /api/slots/{name}) paths.
• Lemonade slots: unaffected — no new fields added.

Frontend

• Profiles page (ui/src/dash/profiles.jsx): ProfilesView component with ProfileCard per profile. Intent labels: moe-rocmfp4 → "MoE agents · ROCmFP4 · ~52.8 tok/s", dense-mtp-rocmfp4 → "Dense chat + MTP · ~24.4 tok/s", vulkan-std → "Vulkan std (fallback)". Image tag + profile name as secondary metadata. Registered in nav, routing, and TopBar.
• useProfiles hook (ui/src/api/hooks/useProfiles.ts): queries GET /api/profiles, 60s stale time.
• Edit drawer container branch (ui/src/dash/slot-modals.jsx):
  • Profile picker replaces Device/Backend selectors for runtime=container slots.
  • n_gpu_layers, rope_freq_base, extra_args are read-only with "defined by profile" hint.
  • idle_timeout_s and workers hidden for container slots.
  • Save body: container slots only send ctx_size + default — explicitly excludes n_gpu_layers, device, rope_freq_base, idle_timeout_s, workers, llamacpp_args (owned by profile).
  • Flags section: container slots show backend-provided resolved_command (real argv); lemond slots keep effectiveFlagsFor() preview unchanged.
  • ctx_size restart warning reworded: "⟳ restarts the container (~model-load seconds)" for container slots.
  • Backend mismatch banner hidden for container slots.
• Create modal: runtime selector (lemonade/container); profile picker for container runtime; creates with { runtime: "container", profile }.

Decision notes

• effectiveFlagsFor() is kept and used unchanged for lemond slots — only container slots use resolved_command. The issue spec says "DELETE effectiveFlagsFor()" but also "lemond unchanged" — keeping it for lemond is the correct interpretation.
• resolved_command is on the list path (via _container_state_enrichment) because EditSlotDrawer gets its slot from the list payload through useSlots().

Tests

• tests/api/test_slots_container_state.py: 10 tests — 3 new for runtime/profile/image/resolved_command fields including a --model llama-3b assertion that guards the nested-dict bug.
• ui/tests/e2e/specs/profiles-page-v3.spec.ts: 9 Playwright tests for profiles page and container drawer.
• Full API suite: 787 passed, 3 skipped, 0 failed.
• UI build: clean (vite, 132 modules).
• ruff: all branch-changed files pass.

Closes #658
Parent epic: #652

[thinmintdev]

VERDICT: APPROVE-ON-GREEN — merge-ready once the python matrix + γ-suite pass

ui already green; python (3.11/3.12) + γ-suite pending. Solid backend + UI work; lemond parity is preserved (the #668 lesson). Two non-blocking notes below.

SPEC (vs #658): ACs met.

• resolved_command_for_slot() correctly reads the model from cfg["model"]["default"] (nested TOML table), not str(cfg["model"]) — the dict-repr bug is fixed and guarded by test asserting "--model llama-3b" in the joined argv. Image + flags (via resolve_profile_flags) + port all assembled correctly; podman preamble omitted by design (needs root GIDs, not useful for debugging).
• /api/slots emits runtime/profile/image/resolved_command on both list + detail paths for container slots; lemonade slots get none of these (test_lemonade_slot_has_no_runtime_container_fields). Graceful degradation to None on profile-lookup failure / absent profile.
• Profiles page: useProfiles hook + ProfilesView with intent labels (MoE agents · ROCmFP4 · ~52.8 / Dense chat + MTP · ~24.4 / Vulkan std fallback), custom-profile fallback to image tag + MTP. Nav/route/TopBar wired. Graceful loading/error/empty states.
• Container edit drawer: profile picker replaces device/backend; n_gpu_layers/rope_freq_base/extra_args read-only with "defined by profile" hint; idle_timeout_s/workers hidden; provider→image tag, declared/actual→profile+image_status, mismatch banner hidden; resolved_command shown instead of effectiveFlagsFor; ctx_size restart copy reworded. Create modal: runtime selector + profile picker, canSave gated on profile for container.
• Container SAVE body correctly excludes profile-owned knobs: isContainerSave gates out n_gpu_layers/device/rope_freq_base/idle_timeout_s/workers/llamacpp_args — only ctx_size + default sent.

STANDARDS — lemond parity PRESERVED (the #668 lesson):

• Every container change in EditSlotDrawer/CreateSlotModal is gated on slot.runtime==="container"; the lemond else branch contains the ORIGINAL code verbatim (Device row, Runtime Backend IIFE, idle/workers rows, extra_args hint, effectiveFlagsFor preview).
• Lemond save body is byte-equivalent: !isContainerSave && X → X, and the ternary spreads resolve to the original {device}/{n_gpu_layers} for lemond. Verified.
• Create modal: runtime defaults to "lemonade", so an operator who ignores the new Runtime dropdown gets identical behavior (device selector + device in body + original rocmfp4 model filter). The added Runtime dropdown is additive, not a regression.
• effectiveFlagsFor() correctly KEPT for lemond (issue AC "lemond keeps its form" outranks the audits "delete it") — conscious, documented tradeoff: lemond still shows the approximate client-side preview while container shows the real resolved_command.
• No fabricated data; container fields degrade to null/"—" when absent.

NON-BLOCKING NOTES (P2 / maintainers call):

1. resolved_command "--model" token shows the model ID (cfg["model"]["default"], e.g. "llama-3b"), NOT the absolute .gguf path the container actually launches with. The real load path (_render_unit via load_sync) uses _resolve_model_path(model_info) → /mnt/ai-models/llama-3b.gguf, so the displayed "real podman argv" diverges from reality on the model arg. The unit test enshrines the id form (--model llama-3b) while the Playwright fixture hand-writes a full .gguf path — inconsistent, so nothing pins the divergence. Its a recognizable read-only preview, AC arguably met in spirit; decide id-vs-path. (Note: the absent --ctx-size is FAITHFUL to #666s _render_unit, which also omits it — not a #671 bug.)
2. Hot read-path cost: _container_state_enrichment now calls load_profiles_config() AND resolved_command_for_slot() (which calls load_profiles_config again) per container slot per /api/slots poll — two TOML disk-reads/parses each, on top of the #667 systemctl+health probes. Cache or pass the catalog through in P2.
3. Minor: except (KeyError, Exception) is redundant (== except Exception); container create hardcodes device:"gpu-rocm" even for a vulkan-std profile (cosmetic — container ignores device, but writes mildly misleading TOML). One-liners.

CONTEXT NOTE: #668 merged with my prior CHANGES open (isSlotLive/stateChipClass-warming parity). #671 branches from main and inherits those but does NOT touch them — out of scope here; they remain a #668 follow-up.

MERGE-READY: yes, on green python + γ. The three notes are P2 polish / a maintainer id-vs-path decision, not blockers. Unblocks #659/#660.