v0.7.54: resource sorting, deployed chat improvements, self hosting docs overhaul, nextjs and bun upgrade - #6237

…ars (#6211)

* fix(tooltip): remove velocity skew/scale that blurred text on every appear

The tooltip animated a fractional scale() + skew() over 150ms on the element
containing its text. Chrome promotes the bubble to a compositor layer for the
transition, rasterizes the text once at the pre-transition scale, then
GPU-resamples that bitmap for the duration — so text rendered blurry until the
transition settled and the layer re-rasterized at 1:1.

It fired on every appear: a pointer entering a trigger is by definition moving,
so the first pointermove after pointerenter always set a non-zero skew and a
fractional scale.

- drop the velocity-reactive skew/scale flourish and the pointer-velocity
  bookkeeping that existed only to feed it
- round tooltip position to whole pixels; clientX/clientY are fractional on
  HiDPI/zoomed displays, leaving the bubble on a subpixel boundary
- drop the dead `filter` from the transition list — nothing ever set a filter
- skip the state update when the rounded position is unchanged, so pointer
  jitter no longer re-renders every Tooltip.Trigger/Content consumer

The 150ms ease-out translate is kept, so the bubble still trails the cursor.

* improvement(tooltip): keep the velocity flourish, drive it without a CSS transition

Restores the velocity-reactive skew/scale removed in the previous commit. The
flourish was never the problem on its own — handing it to a CSS transition was.
An interpolated fractional scale makes the compositor rasterize the tooltip's
text once and resample that bitmap for the duration, which is what read as blur.

Applied as a static value per pointer event instead, so every frame is
rasterized at its own scale:

- split the transform across the individual `translate`, `scale`, and
  `transform: skew()` properties, and transition only `translate` — position
  still eases toward the cursor, the flourish no longer interpolates
- smooth the pointer velocity in JS (low-pass filter) to replace the smoothing
  the CSS transition used to provide, so the squish still ramps rather than
  snapping between raw per-event velocities
- quantize the flourish to 3 decimals so jitter below the visible threshold
  settles instead of re-rendering every consumer

Whole-pixel position rounding and the redundant-update bail-out are unchanged.

* fix(tooltip): don't seed pointer velocity from the trigger box on focus

The previous commit routed `onFocus` through a shared reveal helper that seeds
`lastPointerRef` from the coordinates it is given. For focus those are the
trigger's box center, not the pointer — so if the pointer already happened to be
over the trigger, the next `pointermove` measured the box-to-cursor delta as
velocity and spiked the skew/scale flourish.

Split the helper in two: reveal-from-pointer seeds velocity tracking,
reveal-from-element leaves it cleared. Restores the pre-PR behavior, where focus
explicitly nulled the pointer snapshot.

Caught by Cursor Bugbot.

* fix(tooltip): make the flourish smoothing frame-rate independent

The velocity low-pass filter applied a fixed coefficient per pointer event, so
how fast the squish settled depended on how fast the device emitted events —
233ms at 30Hz down to 29ms at 240Hz, an 8x spread for the same gesture. It was
also far snappier than the 150ms CSS ease-out it replaced, so the flourish read
as twitchier than before.

Derive the coefficient from the real elapsed time instead
(1 - exp(-dt / tau), tau = 50ms). Settling is now flat at ~150ms from 60Hz
upward, matching the duration of the transition this stands in for.

Also separates the smoothing delta from the velocity-normalization delta: the
latter is still floored at one frame to keep a 1ms event from reporting an
enormous velocity, but flooring the former was itself a source of frame-rate
dependence below 16ms.

Verified against Chrome's documented re-raster behavior: a layer is re-rastered
at its new scale when the scale changes via script, but not when a declarative
animation interpolates it, which is why the flourish must stay out of the
transition list.
https://developer.chrome.com/blog/re-rastering-composite

* fix(security): meter and throttle the deployed-chat TTS relay

POST /api/proxy/tts/stream treated "a live public chat exists" as
authorization to spend the platform ElevenLabs key. A public chat id is
handed to every visitor, so any anonymous caller could synthesize speech
with no length cap, no rate limit and no usage accounting.

Bring the relay in line with its STT sibling (/api/speech/token):

- Resolve the chat's workspace and bill synthesized characters to that
  payer via a new `voice-output` usage source, so spend is attributable
  and counts against the plan's usage limit (402 once exceeded).
- Throttle per IP before any database work, and per chat afterwards, to
  bound both one caller hammering many chats and many callers hammering
  one chat.
- Cap `text` at 2000 characters and allowlist `voiceId`/`modelId`, so the
  caller can no longer choose an unbounded charge, a premium or cloned
  voice, or the billing model.
- Drop `Access-Control-Allow-Origin: *`, which let any third-party page
  read the audio; deployed chat and the Office embed are same-origin.

* fix(security): correct TTS metering, pricing and body cap

Follow-up review of the previous commit found five defects in it:

- Usage rows collided. `usage_log.event_key` is unique and inserts are
  conflict-do-nothing, and the key is derived from the entry's stable
  fields. With no explicit sourceReference, two synthesis calls of equal
  character count in the same workspace produced the same key, so every
  repeat length went unbilled — defeating the metering this change is
  for. Each call now carries a unique sourceReference.
- Priced at $0.10 per 1k characters, twice the published ElevenLabs
  Flash/Turbo rate of $0.05, which would have overcharged customers 2x.
- No body cap, so an anonymous caller could make the route buffer up to
  the shared 50 MB default before validation. Now 16 KB, as the STT
  sibling does.
- Threshold settlement ran per sentence: several queries and a possible
  Stripe call on a realtime path. The workflow execution that produced
  the text already settles the payer.
- The per-IP bucket was described as preventing database amplification.
  getClientIp trusts the leftmost X-Forwarded-For, so an attacker rotates
  past it; the comment now says the per-chat bucket is load-bearing.

* refactor(chat): share the deployed-chat auth gate across voice routes

Review of the previous commits surfaced duplication and one more gap:

- The TTS and STT routes had grown near-identical copies of the chat
  auth + payer lookup. Extracted to resolveDeployedChatCaller, so the
  gate and the payer resolve together and cannot drift per route — that
  duplication is how the unmetered TTS path shipped in the first place.
- Neither copy filtered chat.archivedAt, so an archived chat could still
  authorize spend against its former owner's workspace. The shared
  lookup now filters it, fixing both routes at once. Note: not covered
  by a test — the db chain mock does not evaluate WHERE clauses, so an
  assertion here could not fail.
- Replaced the route's hand-rolled 429 builder with the existing
  enforceIpRateLimit helper, and added enforceChatRateLimit alongside
  the per-user/IP/workspace helpers. Gains the standard Retry-After and
  X-RateLimit-Reset headers plus throttle logging.
- Dropped a test that asserted a module the route no longer imports was
  never called: it could not fail.
- Narrowed the contract: unexported the single-use allowlists and
  dropped .passthrough() now that the body is a closed shape.

* fix(security): fail closed when voice-output usage cannot be recorded

Review round 1 findings:

- A ledger write failure previously logged and streamed the audio anyway,
  leaving the spend unrecorded and the payer's usage understated. The
  caller is anonymous, so serving audio we could not charge for is the
  unmetered spend this route exists to prevent — it now returns 500.
- Use generateId() from @sim/utils/id rather than crypto.randomUUID, per
  the AGENTS.md ID rule. generateId returns a full UUID v4, so the
  per-call uniqueness the usage_log event_key depends on is unchanged.

* fix(chat): split long TTS text so the relay cap cannot drop audio

The client sentence-splits on Western `.!?` only, so text that never
matches — CJK punctuation, or a list with no terminal punctuation —
accumulates and is flushed as one block at the end of the stream. Against
the new 2000-character relay cap that block is rejected and the whole
message plays no audio, a regression introduced by adding the cap.

Split to cap-sized pieces at the single point that enqueues synthesis, so
both the per-sentence path and the end-of-stream flush are covered.
Prefers a whitespace or CJK punctuation boundary, falling back to a hard
cut when a block has none. The server cap stays as the enforcement point.

* fix(security): release the vendor stream when metering rejects the request

The fail-closed branch returned 500 with the ElevenLabs response body
still open, so synthesis and download kept consuming vendor and runtime
resources for a caller that was already rejected. Cancel it before
returning, and assert the cancellation in the test.

…6214)

The hosted `<PublicEnvScript>` rendered a plain `<script>`, which lands at
the end of `<head>` — after the ~40 `<script async>` chunk tags Next emits
at the top of the document. An async script runs as soon as its fetch
resolves, so on a warm cache a Next chunk could execute (and hydration
begin) before the parser reached the env tag, leaving `window.__ENV`
undefined for the first render.

That surfaced as "Something went wrong" on the workflow page, since
`getBaseUrl()` throws during the deploy modal's render, and as the socket
falling back to the page origin instead of NEXT_PUBLIC_SOCKET_URL.

Regressed in #5522, which replaced next-runtime-env's PublicEnvScript
(to avoid its unstable_noStore forcing dynamic rendering) with a static
equivalent that dropped the beforeInteractive strategy.

- render the library's own `<EnvScript>`, which defaults to
  beforeInteractive and does not call unstable_noStore — hosted and
  self-hosted now share one implementation and one loading strategy
- drop the hand-rolled serialization and `<` escaping; Next's
  beforeInteractive path already runs the payload through
  htmlEscapeJsonString, which escapes `& > < U+2028 U+2029`
- fall back to the browser origin in getBaseUrl() rather than throwing,
  so a missing injected env can never tear down a page through the error
  boundary; server-side callers still fail loudly
- read NEXT_PUBLIC_EMAIL_PASSWORD_SIGNUP_ENABLED via getEnv() in the SSO
  form, matching login/signup/auth-modal — `env.X` returns the build-time
  placeholder, not the runtime value

Removes the voice-first interface and TTS playback from the deployed
chat, keeping workspace dictation, which is a separate feature.

- Deletes the VoiceInterface UI and its particles canvas, the chat mic
  input, the TTS audio-streaming hook, the /api/proxy/tts/stream relay
  and its contract, and the voice-settings query hook.
- Unpicks the voice wiring in chat.tsx and use-chat-streaming: the
  audio stream handler, sentence-splitting for speech, voice-first mode
  state, and the isVoiceInput plumbing through ChatInput.
- Drops the now-dead chatId branch from /api/speech/token. It was the
  anonymous public-chat path; with no caller left it would have stayed
  an unauthenticated relay spending the platform key. That leaves
  resolveDeployedChatCaller unused, so it goes too.
- Removes code the above orphaned: MAX_CHAT_SESSION_MS, the noop util,
  the audio/position refs in use-chat-streaming that were only ever
  written, and the chatId field on the speech contract.

Keeps /api/settings/voice, lib/speech and use-speech-to-text: the
workspace home input still uses them. Keeps the voice-output usage
source, enum and label — Postgres cannot drop an enum value, and
historical usage_log rows still need a label to render.

Full-file cleanup pass over the files voice mode was removed from.

- /api/speech/token: with the chatId branch gone, billingAttribution is
  always resolved, so four `billingAttribution ? ... : ...` guards and
  the checkActorUsageLimits fallback were unreachable. Removed, along
  with the now-unused imports and test mock.
- chat.tsx: the inputValue state had one remaining writer (`''`) and no
  reader — ChatInput has owned its own input value since it went
  uncontrolled, and removing the voice-transcript caller left
  handleSendMessage always receiving an explicit message.
- use-chat-streaming: messageIdMap was written in three frame handlers
  and never read, and setIsStreamingResponse was returned but never
  destructured by the only consumer.
- Comments: dropped JSX section labels that restate the element beneath
  them and TSDoc that restates the identifier; narrowed the speech
  contract's workspaceId doc, which existed to contrast with the
  removed chatId field.

* improvement(ci): move CodeQL off default setup onto Blacksmith

Default setup pinned every scan to a 4-vCPU GitHub-hosted runner with no
cancel-in-progress: PR scans ran 30-125 min and re-ran on every push (#6183
burned six overlapping runs). None of that is reachable from the settings UI,
so the config moves into the repo.

- Blacksmith 8-vCPU via the same CI_PROVIDER escape hatch as ci.yml
- cancel-in-progress scoped to pull_request so push/schedule scans finish
- push to main + PR to main/staging + nightly safety net
- paths filter so doc-only PRs skip the run entirely
- paths-ignore config drops tests/mocks/fixtures: 12,716 -> 11,128 files
- languages: javascript-typescript + actions; python dropped (7 files in tree)

Default setup has been disabled; the two cannot both be active.

* fix(ci): restore CodeQL coverage of the data-drain test route

Review round 1.

- Drop the '**/test/**' and '**/tests/**' globs. A `test` directory is a
  routable Next.js path segment, and those globs excluded
  apps/sim/app/api/organizations/[id]/data-drains/[drainId]/test/route.ts —
  a POST handler that authorizes, decrypts destination credentials and makes
  an outbound request. CodeQL paths-ignore has no `!` negation to carve it
  back out, and the globs only covered 76 of 12,716 files.
- Add `ready_for_review` to the pull_request activity types. It is not a
  default type, so a PR opened as a draft and later marked ready was skipped
  by the draft guard and never rescanned until the next push.

…6217)

* fix(sso): derive the SSO callback URL during render

`callbackUrl` was seeded into `useState('/workspace')` and overwritten from
a `useEffect` that read `searchParams`, so the first painted frame always
carried the default. On any deep link with `?callbackUrl=`, the "Sign in
with email" and "Sign up" links briefly pointed at `/workspace` instead of
the requested destination, and a click landing in that window navigated to
the wrong place.

- derive `callbackUrl` from `searchParams` during render; the validation
  gate is unchanged, so an off-origin or malformed value still falls back
  to `/workspace`
- keep the warning for a rejected value in an effect, now keyed on the
  param itself rather than the `searchParams` object, so it fires once per
  actual change instead of once per identity change
- add first-frame tests via `renderToString`, which runs no effects and so
  pins exactly the window the old code got wrong

* fix(auth): resolve callback URLs against the app origin server-side

`validateCallbackUrl` compared against a sentinel base
(`https://callback-url-validator.invalid`) when `window` was undefined, so
the server rejected every absolute URL — including the same-origin ones the
function documents as valid. A component deriving a callback URL during
render therefore produced one destination in the SSR markup and a different
one after hydration.

The exposure was not new to the SSO form: `login-form.tsx` and
`signup-form.tsx` already derive their callback URL during render on
`force-dynamic` pages, so both carried the same divergence.

- resolve against the deployment's own origin server-side, so the server
  reaches the same verdict the browser will after hydration
- fall back to the sentinel when the app URL is unset or unparseable, which
  keeps the server fail-closed: absolute URLs are rejected, as before
- cover the absolute same-origin case and the unset-app-URL fallback in the
  existing suite; all 15 open-redirect rejection cases are unchanged

* fix(env): drop the getBaseUrl browser-origin fallback

The fallback was added in #6214 as a safety net while the real cause — the
hosted env script losing its `beforeInteractive` strategy — was fixed in the
same PR. With the injection ordering restored, `window.__ENV` is populated
before hydration, so the fallback is unreachable in any correctly configured
deployment.

Guessing the origin was also unsafe in the one case it could still fire. An
opaque origin — a sandboxed iframe, and `/chat/*` is deliberately embeddable
— serializes to the string `'null'`, which is truthy, so `getBaseUrl()` would
have returned `'null'` and every call site would have silently built
`null/api/...`. A throw surfaces the misconfiguration instead of encoding it
into request URLs.

- restore the unconditional throw when NEXT_PUBLIC_APP_URL is unset or blank
- mirror it back in the shared testing mock
- flip the two fallback tests to assert the throw, keeping whitespace-only
  coverage

Server-side behavior is unchanged: there was never a `window` to fall back
to, so callers that already guard `getBaseUrl()` (`getBaseDomain`,
`validateCallbackUrl`) keep their existing fail-closed paths.

PATCH /api/chat/manage/[id] calls performFullDeploy when the workflow has
drifted from its active deployment, so editing a chat can mint a new
deployment version. useUpdateChat invalidated only chatStatus and
chatDetail, leaving the deployment panel showing the previous version and
a stale "needs redeployment" indicator until the staleTime expired.

Both mutations now route through invalidateDeploymentQueries, the shared
helper the rest of the deployment surface uses. That also picks up
deployedState, which useCreateChat's hand-rolled list had omitted even
though performChatDeploy replaces the deployed workflow state.

Tests cover both mutations and were verified to fail against the previous
invalidation.

* refactor(chat): clean up the deployed chat surface

Eight-angle cleanup pass over the full contents of the chat surface and
the speech code that survived the voice-mode removal.

Dead code
- enforceChatRateLimit: added for the TTS relay in #6212, orphaned when
  #6215 deleted that route. Zero consumers.
- ChatToolCallStatus, ChatErrorType, and six unused CHAT_ERROR_MESSAGES
  keys (only GENERIC_ERROR and CHAT_UNAVAILABLE are read).
- scrollToMessage was declared and destructured by ChatMessageContainer
  but never used in its body; removing the prop also made the
  scrollToShowOnlyMessage branch unreachable, since the sole caller
  passed true.
- permissionState and the language prop on useSpeechToText: both
  write-only across the repo.
- The image branch in ChatFileDownload's renderIcon returned the same
  DefaultFileIcon at the same size as the fallback.
- chatKeys.status/detail: aliases of deploymentKeys nothing imported,
  and misleading since they root under a different key namespace.

Redundant state
- password-auth and email-auth each kept a boolean in lockstep with
  `errors.length > 0`; email-auth also validated on every keystroke and
  then immediately hid the result.
- file-download tracked hover in state to drive one opacity class; now
  group-hover. Verified emcn Button sets no `group` class of its own.

Memoization
- ChatMessageContainer's memo() could never bail: chat.tsx passes an
  inline arrow for scrollToBottom and displayMessages is a fresh array.
  Four of the five things that re-render ChatClient are its props
  anyway, so the memo is dropped rather than propped up.
- ClientChatMessage keeps its memo — it blocks markdown re-parsing —
  but loses the custom comparator, which compared proxies (a
  key:status fingerprint, files by length) and ignored attachments and
  type entirely. Default shallow compare on its single prop is both
  simpler and stricter.
- Six useCallbacks whose consumers are native DOM handlers or inline
  arrows, so nothing observed their identity.

Effects
- The scroll listener attached in an effect keyed on [chatConfig,
  authRequired] — values it never reads, standing in for "the container
  has mounted". It now attaches via a ref callback, so it no longer
  re-attaches on every config refetch.

Design system and a11y
- z-[100] -> z-[var(--z-dropdown)] (same value), shadow-lg ->
  shadow-medium, list styles from inline style to Tailwind classes,
  hover: -> hover-hover: on touch-reachable targets, Check sourced from
  emcn alongside its Duplicate pair.
- Accessible names on the remove-attachment, stop, and send buttons,
  which announced only as "button".
- Dropped a keyboard handler on a role='group' div with no tabIndex,
  where target === currentTarget was unreachable, and the Tooltip
  Provider wrappers and delayDuration, which emcn documents as
  no-op passthroughs.

* fix(chat): restore markdown list markers

The design-system pass swapped inline `listStyleType` for Tailwind
classes, but the edit that added `list-disc`/`list-decimal` silently did
not apply while the one removing the inline style did. With Preflight
setting `list-style: none`, every bullet and number in an assistant
response disappeared. `list-item` on the `li` sets display only, not the
marker type.

…dd self-host settings, land setup on signup (#6216)

* fix(auth): skip email verification when no mail provider is configured

Signup pushed /verify unconditionally, stranding self-hosted deployments
with no mail provider on a screen no email could ever satisfy. Derive one
server-side effective value (verification enabled AND deliverable) and read
it from Better Auth enforcement, signup routing, and the verify page.

* feat(settings): add a self-host section with the managed Chat keys link

Self-hosters had no in-app pointer to the managed service that issues their
Chat keys. New Settings > System > Self-host section, gated on `requiresSelfHosted`
so it is absent on hosted Sim, containing only that link.

* improvement(setup): land the wizard handoff on signup

A freshly provisioned deployment has no accounts and / renders the marketing
landing page, so the bare origin left operators hunting for the CTA. Single-source
the URLs and point every open-Sim handoff at /signup across all three modes.

* improvement(settings): mark the self-host section with a sprout

Server was already doing double duty for MCP servers and Mothership, and the
icon set ships no botanical glyph, so the mark is a text emoji.

* improvement(settings): draw the sprout as an emcn line icon, move to Platform

The emoji rendered in the platform's own colors, so it was the one glyph in the
nav that ignored --text-icon. Replaced with a hand-drawn emcn Sprout (24 grid,
1.55 stroke, currentColor) matching the house style, renamed the tab to
Self hosting, and regrouped it under Platform — self-hosting is deployment-wide,
not per-workspace. Still self-hosted-only.

* improvement(settings): drop the section header from self hosting

One row does not need a section label, and removing it takes the divider with
it. The body is now the Chat keys row and its managed-keys link, nothing else.

* refactor(voice): load STT availability through React Query

useSpeechToText fetched `/api/settings/voice` inside an effect and stored
the result in useState behind a hand-rolled mountedRef guard: no cache, no
dedupe across mounts, and no AbortSignal, so the response was fetched and
parsed even after unmount. Two simultaneously mounted consumers issued two
requests. It also bypassed hooks/queries/**, which is where every other
server read in the app lives — and it escaped `check:react-query`, whose
audit only covers useQuery/useMutation call sites.

The value is server env read at request time, so it cannot change within a
session; the new hook uses an infinite staleTime and a caller-controlled
`enabled` so clients without the audio APIs never issue the request.

Hydration is unchanged: SSR renders unavailable, and the first client
render still resolves unavailable because `data` is undefined until the
fetch settles. No initialData, deliberately — adding it would break that.

mountedRef stays; it is still load-bearing for the streaming lifecycle.

* test(queries): unmount rendered roots between tests

renderHookWithClient created a React root per test but never tore it
down, so trees stayed mounted with live QueryClient observers until
worker teardown and async notifications could cross test boundaries.

Audited every test in the repo using createRoot: 51 of 53 already
unmount. The two that did not were both mine — voice.test.tsx here and
chats.test.tsx from #6223 — so both are fixed and the pattern is now
uniform.

* fix(voice): let a failed STT probe recover on a later mount

The app QueryClient sets retryOnMount: false and retry: 1, and
refetchOnWindowFocus only refetches stale queries — which an infinite
staleTime never becomes. So one transient failure cached the error for
the life of the client and hid the mic until a full page reload.

The effect this replaced refetched on every run, so retryOnMount: true
restores parity: no refetch after success, a retry per mount after
failure. Test asserts recovery under the app's real query defaults and
fails without the override.

…nd deploy modals (#6226)

* improvement(emcn): share one emails/domains chip input across share and deploy modals

Extracts the emails chip lifecycle out of ChipModalField type='emails' into a
standalone ChipEmailsInput, and points both the file share modal and the deploy
modal's chat tab at it instead of their hand-rolled TagInput wiring.

- add ChipEmailsInput (dedupe, normalize, format gate, paste, per-chip errors)
  with an allowDomains opt-in for bare @domain.tld entries
- share modal and deploy modal chat tab now use it; drop both hand-rolled
  add/remove/validate implementations and the dead emailError state
- move the shared allowlist policy into validateAllowlistEntry
- drop the "Add specific emails or whole domains" hint text
- give OutputSelect a size prop; the deploy modal chat tab uses the 30px chip
  trigger so it lines up with the Title field above it
- drop overflow-y-auto from the chat deploy form, which was promoting overflow-x
  to auto and rendering a stray horizontal scrollbar

* improvement(utils): one email syntax gate, drop the backtracking placeholder regex

Audit follow-ups on the emails chip input.

- move EMAIL_SYNTAX_REGEX and the new @Domain pattern into @sim/utils/string as
  isValidEmailSyntax, so emcn and lib/messaging/email/validation.ts stop keeping
  byte-identical copies of the RFC 5322 regex
- allow single-label domains (@intranet) again — requiring a dot rejected
  entries the old startsWith('@') check accepted, which self-hosted
  deployments use. A lone @ and malformed labels stay rejected
- replace derivePlaceholderWithTags' /^Enter\s+(.+?)s?$/i with string ops;
  CodeQL flagged it as polynomial backtracking (js/redos). Verified identical
  output across the placeholder shapes in use
- forward the emails control's props explicitly instead of underscore-discard
  destructuring, matching ChipModalFileControl in the same file

* test(email): pin the allowlist entry rules

Covers isValidEmailSyntax's allowDomains branch (single-label domains stay
valid, malformed bare domains that the old startsWith('@') check accepted do
not), the 254-character cap, the DNS label limit, and validateAllowlistEntry
waiving address-level policy for bare domains. Verified both new rules fail
when the behavior is reverted.

…ting docs (#6225)

* feat(self-host): align Docker Compose with Helm and overhaul self-hosting docs

Docker Compose shipped no scheduler, so scheduled workflows, every polling
trigger, connector syncs, the outbox, and data drains silently never ran.
Adds a cron service running the same 18 jobs the Helm chart schedules as
CronJobs, and closes the remaining behavioral gaps between the two paths:
bundled Redis in the chart, no hosted plan caps in chart defaults, pinned
image tags, and fail-fast secrets. A CI check keeps the schedulers in sync.

Also rewrites the self-hosting docs: 14 new pages, 8 updated, reorganized
into Install / Configure / Operate.

* fix(self-host): drop bun install from chart CI, remove air-gapped and backup docs

The scheduler-parity check pulled a full dependency install into the
chart-validation job, which fails building isolated-vm on that runner.
Rewritten to use only node builtins so the job installs nothing.

Also removes the air-gapped and backup/restore pages, and stops pinning a
concrete release in the docs so the examples do not go stale each release.

* fix(helm): bundle Redis in secret-manager modes unless the URL is supplied

Suppressing Redis whenever a secret mode was active left those deployments
with no Redis at all — REDIS_URL is optional there and both shipped examples
omit it. The chart now steps aside only on a detectable signal: an explicit
app.env.REDIS_URL, an ESO remoteRefs.app.REDIS_URL mapping, or the new
redis.provideUrl=false opt-out for a pre-created Secret it cannot read.

* fix(compose): derive realtime BETTER_AUTH_URL from NEXT_PUBLIC_APP_URL

realtime read BETTER_AUTH_URL directly and fell back to localhost while
simstudio derived it from NEXT_PUBLIC_APP_URL, so setting only the public
origin left realtime authenticating against http://localhost:3000.

* fix(helm): deliver bundled REDIS_URL via ConfigMap so an operator value always wins

Injecting REDIS_URL as an inline container env made it beat every envFrom
source, so a REDIS_URL held in a pre-created Secret or synced by External
Secrets was silently shadowed and traffic moved to a fresh in-cluster Redis.

Kubernetes resolves duplicate envFrom keys by letting the last source win, so
the bundled URL now ships as a ConfigMap listed before the app Secret. Any
operator-supplied value overrides it without the chart needing to read it,
which also removes the redis.provideUrl flag the previous attempt required.

* docs(helm): spell out the egress rule external datastores need

The default NetworkPolicy allows 443 plus the bundled Postgres and Redis by
pod selector. Anything you run outside the chart on another port needs its own
rule, which is easiest to miss when REDIS_URL arrives via a Secret the chart
cannot inspect. Adds a copyable example to the production checklist and the
security guide.

* feat(helm): add networkPolicy.allowExternalEgress for managed datastores

The default policy allows 443 plus the bundled Postgres and Redis by pod
selector, so a managed datastore on another port needs a hand-written CIDR
rule — awkward when REDIS_URL arrives via a Secret the chart cannot inspect.

Adds an opt-in switch that drops the port restriction while still blocking the
cloud metadata endpoints. Defaults to false, keeping this chart stricter than
the common chart default of unrestricted egress.

…terminal features (#6196)

* icon styling

* feat(desktop): isolate chat browser and terminal sessions

* feat(desktop): uncap browser and terminal tabs

* feat(desktop): polish browser and terminal resources

* improvement desktop

* fixes

* updates

* fixes

* fix

* update tests

…6229)

The Redis byte-budget branch in doFlush requeued the rejected batch and
rethrew, skipping the MAX_PENDING_EVENTS trim every other failure path
applies. The backlog then grew for the rest of the run and each retry
re-serialized it, so a wide parallel fan-out could drive unbounded heap
growth and stall the event loop.

Drop rejected chunks instead of requeueing, pace retries through the
existing backoff, and split batches that exceed the single-write cap so
an oversized batch can make progress instead of stalling forever.
Terminal status is now writer-scoped, since a concurrent scheduled flush
can be the loop that drains the final chunk, and a terminal event whose
batch was dropped is retried on its own rather than lost with it.

Record terminal stream meta when the terminal event cannot be buffered,
so reconnecting readers stop polling an active stream until their
deadline. Drop the unused reserve/release budget helpers.

…6231)

* fix(execution): offload buffered event values under budget pressure

An execution buffers EVENT_LIMIT events inside a per-execution byte budget, so
a full ring only fits if events average under budget/EVENT_LIMIT. Values were
only offloaded to object storage at the shared 8 MiB cap, far above that, so a
run emitting large block outputs exhausted its budget within a few dozen events
and stayed pinned at its ceiling for the rest of its life.

Applying that ceiling to every run would be worse than the problem: the SSE
stream carries the compacted event and the terminal renders a ref only as a
preview, so ordinary block outputs would stop being readable live, and every
value would cost an object-storage write on the hot path. Engage the tight
ceiling only once a run has actually buffered past half its budget. A short run
keeps full-fidelity output and pays nothing; a runaway one stops accumulating.

Both bounds derive from the existing budget rather than being asserted, and
preserved UserFile base64 is exempt — it is an explicit request for inline
delivery, already bounded by its own cap and the strip-and-recompact fallback.

Also stop a failed resume-path buffer write from failing the run: it was
awaited bare, so the rejection propagated into the executor callback and failed
work that had already completed. The buffer only backs reconnect replay, so
degrade to live-only delivery the way the execute route does.

* fix(execution): measure pressure at write time and keep terminal status last

Pressure was read from bytes counted once a flush succeeded, but a burst is
compacted long before the scheduled flush runs — so the very batch that
exhausts the budget went through at the loose ceiling and was dropped instead
of offloaded. Count bytes as each event is compacted.

Separately, the terminal-alone retry stamped terminal status while entries
queued ahead of it were still unwritten. Terminal status is the reader's
end-of-run signal: a reconnecting client drains what is in Redis and closes, so
those entries were stranded behind a stream it had already finished with. Drain
the backlog first, then publish the terminal event.

* fix(execution): do not lose the backlog or publish terminal status early

Draining the backlog ahead of the terminal event left the terminal status armed,
so whichever chunk emptied the queue stamped the run complete before its
terminal event was written — the inverse of the ordering the drain was added to
guarantee. Disarm the status for the drain and restore it afterwards.

The drain's result was also discarded: a transient Redis failure requeues its
batch, and the unconditional reassignment that followed dropped those events
even though the budget never rejected them. Keep whatever could not be
persisted, and publish the terminal event alone only once nothing earlier is
still queued — failing otherwise lets the caller degrade, which records the
status without claiming the missing events arrived.

Leave eventId unset on a failed resume-path write. Assigning 0 was persisted by
clients as a reconnect cursor and rewound them to the start of the run.

* fix(execution): keep an event in the buffer when a pressure offload fails

Durable compaction runs before an event is queued, so a storage or metadata
failure dropped it from replay entirely — a reconnecting client would never see
it, even though the live path carried on. Offloading under pressure is only an
optimization that keeps a heavy run from exhausting its budget, so when the
value cannot be persisted, fall back to buffering it inline: exactly what the
run would have done before pressure engaged.

…hain (#6235)

Bumps next, @next/env and the @next/swc-* optional deps to 16.3.0 across the
root overrides, apps/sim, apps/docs and packages/emcn.

Two config notes worth keeping:

- `experimental.turbopackFileSystemCacheForBuild`'s default flipped false ->
  true for stable in 16.3.0, so our explicit `false` is now load-bearing rather
  than defensive. Without it this bump would have silently re-enabled a build
  cache measured 3.2x slower on this codebase (#6078). Comment updated to say so.
- `experimental.useTypeScriptCli: true` is now pinned. TypeScript 7 ships no
  JavaScript compiler API until 7.1, so Next's default checker cannot run and
  needs the project-local `tsc` CLI instead. 16.2.12 was silently skipping
  build-time type checking entirely because it detected @typescript/native-preview
  and short-circuited the stage ("Finished TypeScript in 138ms"); pinning the flag
  keeps that from drifting back.

TypeScript toolchain cleanup that the upgrade makes possible:

- Drop @typescript/native-preview from apps/sim and apps/docs. It was the
  pre-release channel for TS 7 and is superseded by typescript@7 (nightlies now
  ship as typescript@next), and its presence is what suppressed build type checks.
- Align packages/browser-protocol and packages/terminal-protocol from
  typescript ^5.7.3 to ^7.0.2 so every workspace is on one compiler version.

apps/sim keeps @typescript/typescript6 as a production dependency: the function
sandbox at app/api/function/execute dynamically imports the TS 6 compiler API to
transpile user code, and TS 7 has no API to replace it yet.

Build and dev were benchmarked 3x per version on a byte-identical tree; the
upgrade is performance-neutral (build median 100s -> 99s, dev:full warm 10s -> 9s).

…ath (#6232)

* fix(providers): route 6-10MB attachments to the provider large-file path

The inline base64 cap was 10 MB of raw bytes, but the execution payload store
refuses a single value above 8 MiB and base64 inflates by 4/3. Every raw file
over 6 MiB therefore produced a base64 string the store rejected — and the
rejection came from the base64 *cache* write, which threw and failed the run
with "Execution memory limit exceeded" even though the bytes had already been
read successfully. Because shouldUseLargeFilePath only fires above the inline
cap, 6-10 MB attachments had no path at all on any provider: they never reached
the OpenAI or Gemini Files API upload they were supposed to take.

Derive the cap from the payload-store ceiling instead of hardcoding it, and
degrade a refused cache write to "not cached" rather than failing a request
whose bytes are in hand. Every other size guard in the chain compares raw bytes
against maxBytes; only the Redis write sees the encoded size, which is why this
went unnoticed — and why it failed only where Redis is configured.

Also correct the provider ceilings against the vendors' current documentation:

- openai: 50 MiB -> 50,000,000. The gate is `size > maxBytes`, so 50 MiB
  admitted 52,428,800 bytes; the docs say each file must be *under* 50 MB.
- bedrock: had no entry and inherited the inline cap, which is above what
  Converse accepts (3.75 MB per image, 4.5 MB per document).
- groq: 20 MiB -> 20,000,000, and modelled as the request cap the docs
  actually describe rather than a per-file MiB ceiling.
- fireworks: had no entry; its 10 MB budget is on the base64 total, so the
  raw-byte equivalent is 7.5 MB.

Add perRequestMaxBytes for the combined ceilings, enforced before any upload
spend, and cover the OpenAI upload path end to end — it had no test at all.

* chore(deps): upgrade @google/genai to 2.13.0 and @anthropic-ai/sdk to 0.115.0

@google/genai 2.x reworks the Interactions API, which the Gemini deep-research
provider is built on. Migrate it:

- `Interaction.outputs` (a flat content array) is now `steps`, a discriminated
  timeline; the report text lives in the `model_output` steps' text content,
  alongside thought and tool steps we skip.
- `Usage.total_reasoning_tokens` is now `total_thought_tokens`. The old code
  already fell back to that name through a cast, so this just makes the field
  the SDK actually returns the typed one.
- SSE events renamed: `content.delta` -> `step.delta`, `interaction.start` ->
  `interaction.created`, `interaction.complete` -> `interaction.completed`.
  The new event types are discriminated, so the payload casts are gone.

Both `interactions.create` calls also stop annotating their params with
`Interactions.CreateAgentInteractionParams{,Non}Streaming`. In 2.13.0 those
namespace aliases resolve to `CreateAgentInteraction`, whose `stream` is a
plain `boolean` rather than a literal — annotating with them erases the
discriminant and the call resolves to the union-returning overload, so the
result is typed as `Interaction | Stream` at every use. An inline
`stream: true as const` keeps the correct overload.

Neither upgrade required a `minimum-release-age` waiver: 2.15.0 and 0.115.0
were checked and 2.13.0 is the newest genai release clearing the 7-day window.

* chore(deps): upgrade openai to 7.0.0

v5 is the only major with real breaking changes for us; v6 widened a Responses
output type and v7 only raised the Node floor to 22, which apps/sim already
requires. Three things needed fixing:

`ChatCompletionMessageToolCall` became a union of function and custom tool
calls, and the custom variant has no `function` field — 43 unguarded `.function`
accesses across the OpenAI-compatible providers. Narrow once at each
`message.tool_calls` read through a shared `isFunctionToolCall` guard rather
than casting at every use.

That guard deliberately tests for the `function` payload instead of
`type === 'function'`. Many OpenAI-compatible vendors omit `type` on tool calls
entirely — our own fixtures do — so discriminating on it type-checks perfectly
and then silently drops every tool call those providers return.

`ChatCompletionCreateParams.verbosity` narrowed from `string` to a literal
union, and the Responses API's output and input item unions now diverge on
members Sim never emits (computer-use call outputs, whose `status` admits
`failed`, and the `AdditionalTools` escape hatch). Echoing output back as input
is what a tool loop is supposed to do, so that conversion is asserted once in
convertResponseOutputToInputItems and the streaming loop now routes through it
instead of pushing raw output items.

The hand-rolled multipart upload in file-attachments.server.ts can now be
replaced with the SDK's typed `expires_after` — left for a follow-up so this
commit stays a pure upgrade.

* fix(providers): correct defects found auditing the attachment and SDK changes

The mechanical rewrite that added `isFunctionToolCall` to every `tool_calls`
read also rewrote three truthiness guards, where the filtered array was
computed, discarded, and the unfiltered value used in the body. Filter once and
use that value. The helper also landed between `trackForcedToolUsage`'s TSDoc
block and its declaration, leaving that block documenting the wrong function.

Raise the Bedrock ceiling from 3.75 MB to 4.5 MB. Converse caps an image at
3.75 MB and a document at 4.5 MB, and a single `maxBytes` cannot express both.
Taking the lower bound looked conservative but regressed 3.75-4.5 MB documents,
which Converse accepts and which work today. At the document bound every size
that works now still works, and only genuinely-too-large files are rejected
early; oversized images in that band keep surfacing as a Bedrock API error,
exactly as they do without the entry.

Both limits re-verified verbatim against the primary docs: Converse's Message
reference ("Each image's size ... no more than 3.75 MB", "Each document's size
must be no more than 4.5 MB") and Fireworks' vision guide ("Total base64-encoded
images must be less than 10MB").

* fix(providers): keep every attachment that works today working

Auditing the routing change for backwards compatibility turned up two bands it
silently broke.

Lowering the single inline cap made the upload path mandatory above ~6 MB, but
every large-file path reads its bytes back out of cloud object storage. A
deployment without it — local dev, any disk-backed self-host — inlines those
files as base64 today and would have started failing outright with "requires
cloud file storage". Split the one number in two: the inline ceiling stays at
10 MiB, and a separate threshold marks where an upload becomes *preferable*
because the base64 copy no longer fits the payload store. Where no upload path
is reachable, base64 hydration now runs to the inline ceiling as before, and a
missing cloud-storage backend leaves the file for the inline path instead of
throwing.

The two strategies also cross over at different sizes now. `files-api` carries
every type the provider already accepts, so it takes over at the lower
threshold. `remote-url` only fetches images and PDFs, so switching early would
have started rejecting 6-10 MB text documents that inline fine today; it takes
over only once inlining is genuinely impossible.

Revert the Groq ceilings. Its published "20MB" governs a request carrying an
image URL, and on this path the body holds only the URL, so it cannot bind on
the files maxBytes guards. Groq documents no limit on the image it fetches, so
tightening the per-file cap to 20,000,000 and summing raw bytes against the
request cap would both reject uploads that work today on no documented basis.

* fix(providers): drop the provider ceiling changes and close the audit findings

A six-agent line-by-line audit against the vendors' live docs found the
`models.ts` ceiling work was not the strict improvement it was written as, so
all of it is reverted:

- bedrock's 4.5 MB cap broke video. Converse takes image, document AND video
  blocks, and video is allowed 25 MB base64 — a single `maxBytes` cannot
  express three content classes, and every 4.5-10 MiB `.mp4` that works today
  would have started failing.
- openai's combined 50 MB cap is the FILE-input limit. Image inputs are
  governed separately at 512 MB / 1500 images, so summing every attachment
  rejected eight 8 MB PNGs that OpenAI documents as legal.
- fireworks' per-file ceiling was unreachable behind the request budget, while
  the upload picker went on advertising it — a size the UI accepts and
  execution always rejects.
- The whole `perRequestMaxBytes` feature goes with them: it summed raw bytes
  against caps that are variously on encoded bytes, on one content class, or on
  a body that carries only URLs, and it double-counted a file referenced from
  several messages even though the uploader dedupes by key.

Only openai's per-file `maxBytes` stays corrected, to decimal 50,000,000 — the
one number a vendor states unambiguously and writes no MiB against.

Also fixed, all found by the same audit:

The hydration cap stopped short of where `remote-url` actually switches over,
so 6-10 MiB attachments on anthropic/openrouter/xai/groq/together/baseten/vllm
had neither base64 nor a handle and failed outright — the very band this branch
exists to fix. Both decisions now come from one function so they cannot drift.

Eight more sites where the mechanical rewrite computed a filtered array and
then read the unfiltered one (deepseek, sakana, nvidia, kimi), leaving those
providers without the narrowing they appear to have.

`isFunctionToolCall` threw on a null or primitive `tool_calls` entry, because
`in` requires an object — reachable exactly on the self-hosted gateways this
filter was added for. It is now total, and all 32 test mocks match it rather
than being quietly more permissive.

`checkForForcedToolUsage` in utils/litellm/mistral evaluated the response before
the `tool_choice` test, turning a tolerated malformed body into a TypeError on a
path that never used to touch it.

Gemini: `satisfies` restores the excess-property checking the dropped
annotations removed, the poll loop recognises the terminal statuses v2 added
instead of spinning for an hour and reporting a timeout, and the streaming doc
block no longer names five events that were renamed six lines below it.

* fix(providers): report attachment limits in the unit vendors publish

The size ceilings are decimal MB — that is how OpenAI, AWS and Fireworks all
write them — but the error messages divided by 1024², so OpenAI's 50 MB cap
was reported to the user as "48MB". Someone shrinking a 49 MB file to get
under it was chasing a limit that does not exist.

One formatter, used by all three messages, so the file size and the ceiling in
the same sentence are always in the same unit.

* fix(providers): derive the limit unit from the ceiling it belongs to

The previous commit fixed OpenAI's "48MB" by dividing every ceiling by 10⁶ —
which broke the other seven. Only OpenAI's constant is decimal; anthropic,
google, together and openrouter are 50 MiB, baseten and vllm 25 MiB, groq and
xai 20 MiB. Rendering those as decimal MB overstated each by ~5%, so a 21 MB
file on groq was rejected with "(21MB) exceeds the 21MB limit" — a sentence
that contradicts itself and sends the user to shrink a file to a size that is
still over. Same class of bug as the one being fixed, sign flipped.

Both figures now render through one unit taken from the ceiling, so the number
a user is told is the number the vendor publishes and the two sizes in a
sentence are always comparable. Tested against every ceiling in the registry
rather than only the values that happened to round cleanly.

Two more from the same audit:

A file with a missing or zero declared size was stranded on a files-api
provider: hydration bailed on the real byte length while `shouldUseLargeFilePath`
saw `0 > threshold` as false, so it got neither base64 nor a handle and failed
as "may no longer be accessible" — a size failure wearing an access failure's
message. Uploads read the real bytes and enforce the ceiling themselves, so an
unknown size now routes to one.

The oversized-attachment error blamed the provider for a deployment problem:
a files-api provider on a host without cloud storage reported that the provider
"has no large-file upload path", which is not true of the provider.

`isFunctionToolCall` only proves `function` is present, never that it is well
formed, so the trace enricher is defensive again about a hollow payload without
giving up the compile-time gate. The 32 test mocks now match production exactly.

* fix(providers): stop an over-limit size rendering as the limit itself

Deriving the unit from the ceiling fixed the 5% error but left the precision
fixed at two decimals, so a file one byte over a 20 MiB cap still printed
"(20.00MB) exceeds the 20MB agent attachment limit" — the same self-contradicting
sentence, now in a ~5 KB band above every ceiling in the registry. The size
rounds up and the ceiling rounds down, so the two can no longer collide.

The test that was supposed to guard this asserted a file 0.03MB over and an
OpenAI file that was under the limit — neither anywhere near the band — so it
passed while the bug was live. It now walks `limit + 1` for every ceiling, and
goes red against the old rounding.

The reason clause added last commit also claimed a deployment had no cloud file
storage whenever the strategy was not inline. A generated document on a
remote-url provider reaches that same error with storage fully configured,
because a signed URL points at the generation source rather than the rendered
artifact — so it was told something false about its own deployment. That case
now names itself.

* fix(providers): order the attachment failure reason by how general the cause is

The generated-document arm was checked first, so it won over both other causes
and told users two things that were not true.

On an inline-strategy provider — bedrock, mistral, ollama, fireworks, litellm,
vertex, kimi — there is no upload path for any file, generated or not, but the
message blamed the document format and implied a plain PDF would go through.

On openai or google with cloud storage unconfigured it was simply false: a
generated document does take the Files API path there, and that exact file
uploads fine once storage exists. The one actionable fix was hidden from the
operator.

A provider with no upload path cannot be helped by changing the file, and a
deployment with no object storage cannot reach any upload path whatever the
file is, so both now outrank the format-specific case — which is left saying
only what is true of it: a signed URL points at the generation source rather
than the rendered file.

The formatter is unchanged. It was brute-forced over every real ceiling and
three million random pairs with no collision or inversion, but the test's six
ceilings all divide to exact integers, so floor, round and ceil are
indistinguishable on them and the limit-side rounding was unpinned. A ceiling
with a fractional remainder now covers it.

…rbosity, and thinking level (#6233)

* improvement(agent): allow variable references in reasoning effort and verbosity

Reasoning Effort and Verbosity were select-only dropdowns, so a workflow could
not sweep them from a variable or an upstream block the way it already can with
the model. Both become editable comboboxes, matching the model field directly
above them and the managed-agent selectors.

- switch both subblocks to `combobox`, keeping their fetched per-model option
  lists intact
- keep them visible when `model` itself holds a reference, since the concrete
  model id is only known at execution time and cannot be matched against the
  static capability list
- normalize the resolved level in the provider chokepoint so a reference that
  resolves to `"High"` or to nothing behaves sanely instead of hitting a
  provider 400

* improvement(agent): allow variable references in thinking level

Extends the same treatment to Thinking Level so all three model-tuning fields
behave consistently, and logs a level a model does not declare.

- switch `thinkingLevel` to `combobox` with the reference-aware condition
- normalize it alongside the other two; an empty resolve now takes the
  deliberate "send nothing" path rather than the incoherent half-state it hit
  before, and stays distinct from an explicit `none`
- warn when a level is not one the model declares, still forwarding it: Sim's
  per-model lists drive the pickers and can lag a provider, and a sweep needs
  the provider's own error rather than a silent fallback to the default

* improvement(agent): report a model level the sanitizer discards

A model bound to a variable or block reference only resolves at execution time,
so a run whose reference landed on a model outside Sim's catalogue had its
requested level cleared with no signal and quietly fell back to that model's
default. Dropping stays the safe default — a provider with no such parameter
rejects the whole request — but it is now reported.

- log the field, model, and value whenever an unsupported-field level is cleared
- cover both diagnostics, including that they stay quiet for a declared level
  and for the `auto` / `none` sentinels

* fix(providers): redact resolved level content from sanitizer diagnostics

The model-level fields accept environment and block references, so an
unrecognized level is not necessarily a mistyped level — it is whatever the
reference resolved to, which can be secret content. The diagnostics added for
dropped and undeclared levels echoed it straight into server logs.

- log a level only when the catalogue declares it somewhere, or it is an
  `auto` / `none` sentinel; anything else is reported by length alone
- stop discarding levels for a model the catalogue has never seen. Absent is
  unknown, not known-incapable, and a reference is exactly how a newly released
  model arrives before Sim catalogues it — the provider decides instead. Models
  the catalogue knows, and every dynamic-provider id, keep the protective drop

* fix(providers): redact the level in Anthropic's unsupported-thinking warning

Forwarding an undeclared level is deliberate, but it means the Anthropic adapter
receives it and interpolates it straight into its "not supported, ignoring"
warning. Since the field is reference-bound, that value can be whatever a
mistyped `{{ENV_VAR}}` or block reference resolved to — so the redaction added
for the sanitizer's own diagnostics was leaking one layer downstream.

- promote the level renderer to `providers/utils` as `describeModelLevel`, the
  single gate every site echoing a caller-supplied level goes through
- use it in Anthropic's warning and in both sanitizer diagnostics

* refactor(providers): drop the sanitizer's level diagnostics

The two warnings logged server-side, where the workflow author who set the level
never sees them, and the surprising case they described — a level discarded for
a model newer than the catalogue — is now fixed at the source rather than
narrated. They also carried the redaction that leaked resolved content before it
was caught, so removing them removes that surface entirely.

Levels still normalize, and still drop for a catalogued model that does not take
the field. `describeModelLevel` stays for Anthropic's unsupported-thinking
warning, which is a pre-existing log this feature newly exposes to resolved
reference content.

* fix(docker): upgrade bun to 1.3.14 to unbreak the Next 16.3.0 server

Bun 1.3.13 cannot load Next 16.3.0's compiled server runtime. The app container
runs the Next server under Bun (`oven/bun:1.3.13-slim`, `bun apps/sim/bootstrap.js`),
so every app-page render threw and `/api/health` returned 500:

  ⨯ Error: Failed to load external module
    next/dist/compiled/next-server/app-page-turbo.runtime.prod.js:
    TypeError: Expected CommonJS module to have a function wrapper.
    If you weren't messing around with Bun's internals, this is a bug in Bun

Isolated to Bun, not Next, by loading that exact module in the real images:

  Next 16.2.12 + Bun 1.3.13 -> loads (why staging was fine before)
  Next 16.3.0  + Bun 1.3.13 -> CJS wrapper error
  Next 16.3.0  + Bun 1.3.14 -> loads

Bun 1.3.14 is the current stable and already fixes it, so this bumps every pin
rather than reverting the framework upgrade, which would only defer the same
latent Bun bug to the next attempt.

Why no gate caught it: local dev machines and this bump's own verification run
Bun 1.3.14, while the container and CI pinned 1.3.13 — and CI only *builds* the
image, it never boots one and probes `/api/health`. A container smoke test in CI
would have caught this before merge; that is worth adding separately.

* fix(docker): align the remaining bun pins with 1.3.14

Two pins were missed in the first pass because the search was scoped to
docker/, package.json and .github/workflows/:

- .devcontainer/Dockerfile still built on oven/bun:1.3.13-alpine
- PI_BUN_VERSION in apps/sim/scripts/pi-sandbox-packages.ts was still 1.3.13,
  despite being documented as mirroring the root packageManager field, so Pi
  sandbox images would have kept installing the Bun release that cannot load
  the Next 16.3.0 server runtime.

Fixed surgically rather than with a repo-wide replace: "1.3.13" also appears
inside SVG path data in apps/sim/components/icons.tsx and
apps/docs/components/icons.tsx, which a blind sed would have corrupted.