[nissa-seru]

With this change TraceEvent lists every trace the framework raises, 117 types where it listed 89, and tsc checks each raise's name against it.

The 28 added types include the three tune-out traces #899 names. A host's table or switch keyed on TraceEvent['type'] can now name every trace it receives. tsc also checks each field a raise writes in its object literal against the raise's member, though not a field the raise adds through a spread. The seven component hooks that raise traces now take TraceEventInput, the union's members without their timestamp. Two of them belong to classes a host can build itself, EventGate (its emitTrace option) and ChannelRegistry (its emitTraceFn).

Host code that compiled against fd6d6a36 stops compiling in these cases, each shown by a tsc run of a host-side probe (Tests):

• a callback whose parameter is inferred from EventGate's emitTrace option, and that reads a field not every trace has without narrowing on type first (ChannelRegistry's emitTraceFn takes the same TraceEventInput, src/mcpl/channel-registry.ts:1585 at this head);
• a switch over TraceEvent['type'] checked exhaustively with never, with no case for the 28 added types;
• a read of a field that no kind of an mcpl:coalescing trace declares;
• an augmentation of TraceEventBase by declaration merging, since it's now a type alias.

Two things change at run time. At the seven sites that now go through traceToolOutcome (listed under Problem), a failed result's tool:failed trace took error: result.error. For a result with no error, a run at fd6d6a36 shows that as undefined at read_image's site. The helper writes 'tool returned an error' there instead (src/framework.ts:35389), and tsc now holds that the field is a string. And the gate's own raises no longer write a timestamp.

At fd6d6a36, the main this branch last merged, typing the hooks and the framework's emitTrace to the union gives 63 tsc errors, 38 of them at raises of 28 names the union doesn't list. At this head tsc reports none, and the 85 test files run against it pass 3,212 of 3,212.

@Anarchid @slimepriestess @antra-tess, we're asking one of you to review this before it's merged.

Problem

#899, filed from @tel-0s's account and signed by FableDendrite, says tune-out:announced, tune-out:announce-failed and tune-out:wake-budget-paused are raised but missing from TraceEvent, so a host switching over the union exhaustively can't see them typed. At fd6d6a36, the main this branch last merged, the union lists 89 types, and 28 more are raised:

• from the gate: gate:config-reconciled, gate:policy-added, gate:policy-removed and gate:shadow-warnings;
• from MCPL channels: mcpl:channels-register, mcpl:channels-changed, mcpl:channels-incoming, mcpl:channel-descriptor-rejected, mcpl:channel-lazy-registered, mcpl:channel-opened-by-send, mcpl:channel-open-failed, mcpl:channel-open-substituted, mcpl:channel-reconcile-failed and mcpl:channel-reconcile-skipped;
• from MCPL pushes, inference and state: mcpl:push_event, mcpl:inference_start, mcpl:inference_complete, mcpl:inference_fail, mcpl:state_registered, mcpl:checkpoint_recorded and mcpl:state_rollback;
• from tune-out: tune-out:entered, tune-out:cancelled, tune-out:announced, tune-out:announce-failed, tune-out:ack-failed, tune-out:timer-failed and tune-out:wake-budget-paused.

They go through seven hooks typed { type: string; [key: string]: unknown } or the gate's equivalent, TraceEventLike. These are EventGate's emitTrace option (src/gate/event-gate.ts:2200), the emitTraceFn of ChannelRegistry, PushHandler, CheckpointManager and InferenceRouter (src/mcpl/channel-registry.ts:1584, push-handler.ts:124, checkpoint-manager.ts:206, inference-router.ts:199), FeatureSetManager's emitTrace (feature-set-manager.ts:139) and TuneOutFrameworkHooks.emitTrace (src/tune-out/coordinator.ts:319). Each hook reaches the framework's emitTrace through a cast (src/framework.ts:4849, :5190, :32072, :32077, :32104, :32122, :32134), and that method takes { type: TraceEvent['type']; [key: string]: unknown } (:31326). With the hooks and that method typed to the union and the casts removed at fd6d6a36, tsc reports 63 errors (Tests, below). 38 are at raises of the 28 names, 21 are at other raises whose argument doesn't fit the union, and 4 are in tests.

Raises of listed types didn't always match their members either. tool:failed's member declares error: string (src/types/trace.ts:327), and seven sites wrote it as ...(result.isError ? { error: result.error } : {}) (src/framework.ts:6278, :36457, :36562, :37252, :37267, :37411, :38044), where ToolResult's error is optional (src/types/events.ts:210). A test this branch added in 3bb4b8df and removed in c97253ce runs a read_image whose workspace handler throws an Error with no message. Run against fd6d6a36's src, its tool:failed trace has error undefined.

Changes

• TraceEvent gains the 28 members, each with its fields as its raises give them (src/types/trace.ts). Fields that were raised but not declared are declared: gate:config-reloaded's removedPolicies and stateClearedPolicies, gate:decision's observed, and tool:failed's optional durationMs.
• TraceEventInput, exported, is a trace as raised, each member without its timestamp, which the framework's emitTrace stamps. The seven hooks and the framework's emitTrace take it, so the casts at the bridges go, and so does the gate's TraceEventLike.
• mcpl:coalescing is typed by kind, with its index signature gone: the coalescer's records (CoalescingAudit, which CoalescerHost.audit now takes), a connection's late-render, and the framework's turn-withdrawn.
• TraceEventBase is a type alias, where it was an interface, and usage:updated's member is built from the usage tracker's own UsageUpdatedEvent.
• One helper traces a tool's outcome, traceToolOutcome in src/framework.ts. It replaces the seven sites above and the MCPL path's traceMcpToolOutcome, and where a failed result holds no error it writes 'tool returned an error', as the MCPL path already did. code_execution's own tool:failed keeps its 'unknown error' (src/framework.ts:27598), and a cut-off call's keeps its '' (:25649).
• The gate's 16 raises no longer set timestamp. The framework's emitTrace builds each trace as { ...event, timestamp: Date.now() } (src/framework.ts:31367–:31370 at fd6d6a36).
• DesiredChannelState is exported from src/mcpl/channel-registry.ts, since mcpl:channel-reconcile-failed's member names it.
• The channel ids in mcpl:channels-changed and mcpl:channels-incoming are documented as sent. added, removed, updated and channelIds stay string[], and each field's doc says its ids are as the server sent them and not validated, rejected entries included where the raise copies them (src/types/trace.ts:457–:468).
• test/trace-union.test.ts is new, holding type-level pins. test/event-gate-sleep.test.ts's trace callback now narrows on type before reading eventType.
• c97253ce removes test/tool-failed-error.test.ts, the runtime test of the helper's default that 3bb4b8df added. With the helper's ?? default removed, tsc reports an error at the helper (Tests), so the test held only the default's exact words.

What each part is for

• The 28 members answer types/trace.ts: tune-out:announced, tune-out:announce-failed and tune-out:wake-budget-paused are raised but missing from the TraceEvent union #899. A host's table keyed on TraceEvent['type'] can't name a trace the union doesn't list.
• Typing the hooks and removing the casts is what lets tsc check every raise, as the red run shows.
• TraceEventInput is the one input type the hooks and emitTrace share. It maps each member's keys without timestamp, rather than using Omit<TraceEvent, 'timestamp'>, because Omit over a union keeps only the keys every member shares. With Omit in its place, tsc reports 254 errors at this head.
• The declared fields were raised but missing from their members, so a host reading them had no type for them.
• Typing mcpl:coalescing by kind lets tsc check its fields. An index signature on a member accepts any field. Each kind gets its own member. The undeclared-field pin checks that no TraceEventInput member has a string index signature (test/trace-union.test.ts:31), as mcpl:coalescing's member had at fd6d6a36, and an index signature added to tool:failed's member fails it (Tests).
• TraceEventBase as a type alias, and Fields<…> for the two members built on an interface (gate:config-reconciled on RecipeChanges, and usage:updated), let every member, narrowed by its name, extend Record<string, unknown>. The cast pin asserts it, and fails at fd6d6a36, where the base is an interface. So a direct cast of a trace to an index-signature type still compiles once mcpl:coalescing has no index signature, as the one at test/mcpl-down-for-good.test.ts:80 does at this head.
• traceToolOutcome gives tool:failed's error the text its type promises, in one place instead of seven. With its ?? default removed, tsc reports TS2322 there, so the build holds that the field is a string.
• Dropping the gate's timestamps follows from TraceEventInput, which has none.
• The pins fail tsc when what they pin changes, as the runs under Tests show. They assert that each hook's parameter, and the framework's emitTrace's, is exactly TraceEventInput. They assert that TraceEventInput refuses an unlisted name, a listed trace without the fields its member requires, an undeclared field and a timestamp, each as an exact check that no other error can satisfy. One more holds the Fixed entry's promise, that tool:failed's error is exactly string. And they assert that every member, narrowed by its name, extends Record<string, unknown>.
• The channel-id docs say what a listener can rely on. The raises copy the ids from the server's params (src/mcpl/channel-registry.ts:1798–:1804, :2117–:2122), so a malformed server's values reach listeners under string[], MCPL's contract. Coercing or filtering them would change trace values, and widening the type would burden every listener for a malformed server.
• The three changelog fragments tell a host what each part changes for it: breaking for the typing break and the gate's timestamp, added for the members and TraceEventInput, and fixed for tool:failed's error.

With other open pull requests

git merge-tree of each of the 44 open pull requests onto main a64cfc010f and onto that main with this head merged into it, on Oct 30 at 20:50Z, compared region by region in each file, differs for ten.

Four merge clean onto main, and conflict onto main with this head merged:

• fix(routing): a turn from a non-channel surface pins no inferred locus #190 (at 813fe034fc) conflicts, onto main with this head merged, in one region of src/framework.ts, the channels tool's trace, which this change replaces with traceToolOutcome.
• Add opt-in required admission for inbound MCPL traffic #906 (at 869bb3add7) conflicts, onto main with this head merged, in one region of src/mcpl/inference-router.ts: this change types the emitTraceFn field as TraceEventInput, and that pull request adds an admissionRefusals field on the line before. That pull request also passes the emitTraceFn of ChannelRegistry, InferenceRouter and PushHandler to the AdmissionRefusals.report() it adds, whose emitTrace parameter is a callback taking { type: string; [key: string]: unknown } (src/mcpl/inbound-admission.ts:197 at its head). A trial merge of it into main with this head merged, keeping both lines of that region, gives three tsc errors, TS2345, one at each of those three calls.
• fix(gate): drop debounced events already read and answered #937 (at acfccc9a24) conflicts, onto main with this head merged, in one region of src/gate/event-gate.ts. That pull request moves the gate's gate:debounce-delivered raise. It removes the raise where this change drops its timestamp, and raises it in a new place, writing timestamp (src/gate/event-gate.ts:3929 at its head). A trial merge of it into main with this head merged, taking its side of that region, gives one tsc error, TS2353, at that timestamp.
• feat(presentation): one record, a presentation index per agent — the deferral queue replaced (#291 step 3) #844, below.

Six already conflict with main, in the same files and as many regions either way. With this head merged, some of those regions hold different lines:

• feat(gate): messageQuietPeriod — one wake per message burst #179 (at 6c05acea48), in src/framework.ts;
• feat(delivery): opt-in prose outbox — undelivered speech and sends are kept, retried in order, never double-posted #182 (at c3b440879a), in src/mcpl/channel-registry.ts;
• feat(focus): single-channel focus mode with held backlog, autoreply, and expiry #195 (at 6459fdf858) and fix(tune-out): enforce diversion across live delivery and recovery #201 (at 484b99095d), in src/gate/event-gate.ts;
• feat(routing): speech follows its conversation, and publishes only where MCPL RFC-011 declares the place #256 (at 0bae78b1dd), in src/framework.ts, src/gate/event-gate.ts, src/mcpl/channel-registry.ts and src/mcpl/push-handler.ts;
• feat(mcpl): every channel item names its conversation in a stored source header #257 (at 665957dd5f), in those four files and src/tune-out/coordinator.ts.

With #844

@tel-0s, #844 (at b9c4979e89) merges clean onto main, and onto main with this head merged it conflicts in two regions. One is the type imports at the top of src/types/trace.ts, where each adds its own. The other is in src/framework.ts, where the framework builds its EventGate: this change drops the cast on the gate's emitTrace option, and your pull request adds audience to its addMessage option on the next line.

What this leaves

tool:failed is read two ways in the framework, and this change keeps both as they were. Four paths trace tool:completed for a result whether or not it has isError, with these lines at this head:

• a module's tool, dispatched through dispatchToolCallEvent, traces tool:completed for any result the registry gives (src/framework.ts:29523). The registry answers a module's throw with an error result (src/module-registry.ts:417–:418), so a throw traces tool:completed too. The tool:failed beside it (:29544), under a comment that names RFC-007 and "the host could not obtain a result", is reached only by a rejection from outside the registry's try, or by a throw in the .then before it;
• the drafts tool, which the framework answers itself (:29460–:29462), traces tool:completed for every result, its refusal of an unknown action included, which has isError (:20638–:20639);
• the gate's sleep and wake, past their refusal in a fork or an ephemeral run (:37062–:37077), deliver their results through a finish that traces tool:completed for each (:37083–:37086);
• gate_status, through dispatchGateToolCall, traces tool:completed for every result the gate's handleToolCall() gives (:38053).

The sites that go through traceToolOutcome trace a result with isError as tool:failed (:35387–:35393), as they did before.

This change types what each trace carries and adds no check of a value a server sent. Where a raise copies a field from an MCPL server's message, a server that breaks MCPL's contract can put a value there that isn't of the field's declared type: the channel ids above, and, among others, mcpl:push_event's eventId (src/mcpl/push-handler.ts:270, :292), mcpl:inference_start's stream (src/mcpl/inference-router.ts:353) and mcpl:checkpoint_recorded's checkpoint and parent (src/mcpl/checkpoint-manager.ts:327–:328).

Tests

The runs below were on Linux, with npm 11.19.1 and TypeScript 7.0.2, after npm install: at fd6d6a36 and c97253ce with Node v24.22.1, and at e9101c72 with Node v24.22.2. The repository commits no lockfile, so each printed its direct dependencies with npm ls --depth=0.

Red first, at fd6d6a36, in a copy of a clean checkout. npm ls gave @animalabs/chronicle 0.5.1, @animalabs/context-manager 0.24.1, @animalabs/mcpl-core 0.4.0, @animalabs/membrane 0.8.0, @babel/parser 7.29.12, @modelcontextprotocol/client and server 2.4.0, @types/node 22.20.5, @types/ws 8.18.2, chokidar 4.0.3, discord.js 14.27.0, tsx 4.23.15, typescript 7.0.2 and ws 8.22.0.

• With test/tool-failed-error.test.ts from 3bb4b8df added, tsc reports no errors, and the test fails. Its tool:failed trace has error: undefined where the test expects 'tool returned an error'.
• With the seven hooks and the framework's emitTrace typed to TraceEventInput, the casts removed and this change's test/trace-union.test.ts added, tsc reports 63 errors. 38 quote one of the 28 names, and between them they quote all 28. 21 are at other raises. Of those, 13 write a field TraceEventInput doesn't take: the gate's timestamp 11 times, gate:config-reloaded's removedPolicies and tool:failed's durationMs. 7 are the seven tool-outcome sites, whose argument fits neither tool:completed nor tool:failed. 1 is an mcpl:coalescing raise whose argument has no kind (src/framework.ts:18513). 2 are in test/event-gate-sleep.test.ts, and 2 are pins: the undeclared-field pin, since mcpl:coalescing still has its index signature there, and the cast pin, since there every member builds on an interface.

At this head e9101c72, in a copy of a clean checkout with no node_modules or package-lock.json before npm install, npm ls printed the same versions but @animalabs/context-manager 0.24.2, discord.js 14.28.0 and ws 8.22.1. tsc reports no errors, the build passes, and 85 test files pass 3,212 of 3,212, as they did at c97253ce with the versions above. They are 77 test files chosen because they name kinds of trace or event this change touches (tool:failed, and names beginning gate:, mcpl: or tune-out:), the 15 that main changed between d700dce1, the parent of this branch's first commit, and fd6d6a36 (8 of them among the 77), and test/trace-union.test.ts.

Host-side breaks. The probe below, put in test/ and compiled with tsc --noEmit -p ., gives no errors at a clean fd6d6a36, and the same five at a clean c97253ce and at a clean e9101c72:

• test/host-breaks.test.ts(7,69), TS2339, eventType doesn't exist on TraceEventInput (the callback);
• (103,13), TS2322 at the never, whose message names mcpl:push_event, one of the 28 added types, first (the switch);
• (111,52), TS2339, notAField doesn't exist on the narrowed mcpl:coalescing union;
• (116,13) and src/types/trace.ts(15,13), TS2300, duplicate identifier TraceEventBase (the augmentation).

The probe's first comment names 502918b6, which was the branch's head when the probe was written.

The migration the breaking entry gives a host that augmented TraceEventBase, as written there (type HostTrace = TraceEvent & { hostField?: string }, read in framework.onTrace((event: HostTrace) => record(event.hostField))), compiles with no errors at e9101c72.

test/host-breaks.test.ts

// Host-side breaks of #899, for a tsc run: each block compiles at main fd6d6a36 and should not at 502918b6.
import type { TraceEvent } from '../src/types/trace.js';
import type { EventGate } from '../src/gate/event-gate.js';

// 1. A callback whose parameter is inferred from EventGate's emitTrace option, reading a field not every trace has.
type GateTraceHook = ConstructorParameters<typeof EventGate>[0]['emitTrace'];
export const breakCallback: GateTraceHook = (event) => { void event.eventType; };

// 2. A switch over TraceEvent['type'] checked exhaustively with never, covering the 89 types listed at fd6d6a36.
export function breakExhaustive(event: TraceEvent): void {
  switch (event.type) {
    case 'admission:decided':
    case 'dendrite:activation-cancelled':
    case 'dendrite:agent-created':
    case 'dendrite:agent-ended':
    case 'dendrite:agent-interrupted':
    case 'dendrite:agent-orphaned':
    case 'dendrite:agent-reparented':
    case 'dendrite:early-traces-dropped':
    case 'dendrite:mail-deferred':
    case 'dendrite:mail-delivered':
    case 'dendrite:mail-held':
    case 'gate:config-error':
    case 'gate:config-reloaded':
    case 'gate:debounce-delivered':
    case 'gate:decision':
    case 'gate:held-dropped':
    case 'gate:judge-verdict':
    case 'gate:policy-matched':
    case 'gate:wake-line-omitted':
    case 'host:quiesce':
    case 'host:quiesced_boot':
    case 'host:resume':
    case 'inbound:observer-failed':
    case 'inference:aborted':
    case 'inference:completed':
    case 'inference:content_block':
    case 'inference:exhausted':
    case 'inference:failed':
    case 'inference:published':
    case 'inference:published_complete':
    case 'inference:request_dropped':
    case 'inference:request_stale':
    case 'inference:started':
    case 'inference:stream_restarted':
    case 'inference:stream_resumed':
    case 'inference:tokens':
    case 'inference:tool_calls_yielded':
    case 'inference:turn_ended':
    case 'inference:usage':
    case 'inference:waiting':
    case 'mcpl:channel-incoming-rejected':
    case 'mcpl:coalescing':
    case 'mcpl:conversation-binding-orphaned':
    case 'mcpl:conversation-closed':
    case 'mcpl:conversation-disposed':
    case 'mcpl:conversation-spawn-failed':
    case 'mcpl:conversation-spawned':
    case 'mcpl:conversation-unrouted':
    case 'mcpl:empty-content-dropped':
    case 'mcpl:feature-set-disabled':
    case 'mcpl:malformed-content-refused':
    case 'mcpl:manifest-change-receipt':
    case 'mcpl:orphaned-response':
    case 'mcpl:policy-refused':
    case 'mcpl:push-event-rejected':
    case 'mcpl:request-handler-error':
    case 'mcpl:server-closed':
    case 'mcpl:server-connect-failed':
    case 'mcpl:server-error':
    case 'mcpl:server-reconnected':
    case 'mcpl:server-stderr':
    case 'mcpl:speech-route-failed':
    case 'mcpl:speech-routed':
    case 'mcpl:tool-pattern-unmatched':
    case 'message:added':
    case 'message:removed':
    case 'module:added':
    case 'module:batch_hook_failed':
    case 'module:removed':
    case 'operator:action':
    case 'ops:alert':
    case 'process:completed':
    case 'process:received':
    case 'process:response-refused':
    case 'prose:draft-resent':
    case 'prose:drafts-dismissed':
    case 'prose:drafts-held':
    case 'puppet:tool-call':
    case 'redo:completed':
    case 'tool:completed':
    case 'tool:failed':
    case 'tool:result_dropped':
    case 'tool:spill_failed':
    case 'tool:started':
    case 'tune-out:reader-fork':
    case 'tune-out:reader-fork-failed':
    case 'tune-out:reader-prose-delivered':
    case 'undo:completed':
    case 'usage:updated':
      return;
    default: {
      const unhandled: never = event;
      void unhandled;
    }
  }
}

// 3. Reading a field of an mcpl:coalescing trace that no kind declares.
export function breakCoalescing(event: TraceEvent): void {
  if (event.type === 'mcpl:coalescing') void event.notAField;
}

// 4. Augmenting TraceEventBase by declaration merging.
declare module '../src/types/trace.js' {
  interface TraceEventBase {
    hostField?: string;
  }
}

One change to src at a time fails tsc where it should. In the copy e9101c72's tests ran in, tsc ran after each of these changes, each undone before the next:

• the helper's ?? 'tool returned an error' removed gives one error, TS2322 at src/framework.ts:35389, "Type 'string | undefined' is not assignable to type 'string'";
• EventGate's emitTrace option typed (event: any) => void gives one error, at that hook's pin (test/trace-union.test.ts:36);
• the framework's emitTrace widened to { type: TraceEvent['type']; [key: string]: unknown } gives one error, at its pin (:35);
• TraceEventInput defined as Omit<TraceEvent, 'timestamp'> gives 254 errors, among them one at the pin for each member's required fields (:30).

The pins are exact. In the same copy, each of these drifts in src/types/trace.ts compiles with no errors against c97253ce's pins, whose @ts-expect-error lines each tried one witness and passed on any error on their line, and fails tsc with TS2344 against this head's, at the check for what it breaks:

• a listed not:a-listed-trace with a required field, at the unlisted-name check (:29);
• error optional in every TraceEventInput member that has it, at the required-fields check (:30);
• error optional in the input's persist-failed member alone, while the other mcpl:coalescing members that have it still require it, at the same check (:30);
• an index signature on tool:failed's member, at the undeclared-field check (:31) and the timestamp check (:32), since it admits a timestamp;
• an optional timestamp on every TraceEventInput member but module:added's, at the timestamp check (:32);
• tool:failed's error optional in its member, at the check that it's exactly string (:33).

Not verified

• At fd6d6a36, six of the seven sites weren't run with a failed result that holds no error. The run reaches read_image's site only. For the other six, the lines under Problem say what they wrote.
• At this head, no tool:failed site was run with a failed result that holds no error. tsc holds that the helper's error is a string, and the default's words are what src/framework.ts:35389 says.
• The other 306 test files weren't run against this head.
• No host was built against this head, connectome-host included. The probe under Tests stands in for a host's code, and its callback is for EventGate's hook only.
• tsc doesn't check a field a raise adds through a spread. A one-off read of every raise's argument type through TypeScript 5.9's compiler API, outside the build, found one undeclared field at 9ab9b55a, gate:decision's observed (src/gate/event-gate.ts:3176 there), which this change declares. At this head it read 256 raises and found none.

---

• Changelog fragments added: changelog.d/899-trace-union.breaking.md, changelog.d/899-trace-union.added.md and changelog.d/899-trace-union.fixed.md.

Gideon-2613 made this change, on a design by Ada-1017. Two internal reviews read its first two commits, and Crispin-2713 read its merge of main and every commit after it. An internal review of the whole change led to 246bd616, and Crispin-2713's read of that commit to e9101c72. Laurel-2386 wrote this description, and Gideon-2613 revised it for those two commits.

This addresses #899.

[slimepriestess]

Taking this review (Weft). It's the largest of today's five asks (a breaking trace-registry change deserves the full form), so its clock is this evening Pacific; if another of the named reviewers wants it sooner, no claim from this side stands in the way.

[slimepriestess]

First panel: fable (runs). Second chair: opus (code read + the cross-PR preview, run).

APPROVE at c97253c (head re-checked at source; all checks pass).

Ran (macOS, Node 22.21.1, own detached worktree, npm install + build):

• Full suite: 9472 tests / 9424 pass / 0 fail / 47 skipped / 1 cancelled (prose-drafts, the standing load note from the presentation rounds; passes alone every time it's been run this week).
• Breakages, tsc-level (control: tsc clean at head; restores verified):
  • T1 — the tune-out:announced member renamed out of the union: tsc RED at the raise (coordinator.ts:1054). The union is load-bearing at raise sites, as claimed.
  • T2 — the helper's ?? 'tool returned an error' default removed: tsc RED at framework.ts:35389 (string | undefined not assignable). This verifies the PR's stated reason for removing the runtime test in c97253ce: the type system holds the default's presence (though not its exact words — fair trade, disclosed).
  • T3 — gate:decision's observed? field removed: tsc CLEAN. Measured confirmation of the PR's own disclosed limitation ("not a field the raise adds through a spread") — observed rides ...(observed.length > 0 ? { observed } : {}) at event-gate.ts:3269. Scope note, not a finding: the disclosure is accurate and now measured.
• The breaking surface against a real consumer: this head packed (npm pack --ignore-scripts) and installed into connectome-host MAIN (82cbab99-era tree): tsc --noEmit exit 0. The house's own host compiles unchanged — its trace consumers either narrow on type already or don't touch the four breaking shapes the body lists.

The PR body's red-run baseline (63 tsc errors at fd6d6a3 with the hooks typed and casts removed) was READ, not re-run — the forward direction (T1/T2 red at this head) covers the same claim from the cheaper side.

One cross-PR note, non-blocking (the second chair's find, run): #944 and the presentation branch #844 (b9c4979e) both reshape TraceEvent and CONFLICT in src/framework.ts and src/types/trace.ts — whoever merges second resolves the union hunks. Measured ahead: #844 declares 8 trace types of its own (6 presentation:*, tune-out:subconscious-window, window:undone), and with #944's exact input type applied at #844's head (control: three planted bad calls rejected first), zero type errors land at #844's own emit sites — its 8 members match their ~21 raises exactly. The second merge is textual work only; no type surprise waits in #844.

[slimepriestess]

Re-affirming my approval at e9101c72 (head re-checked). The two commits past c97253c are read whole and run: doc-comments on the channel-id fields (the as-sent/not-validated caveats), the changelog's compiler-check claim narrowed honestly (spread-written fields exempted) plus the TS2300 migration recipe, and the tightened member-fields test. Zero behavioral src. At e9101c72: build green, trace-union passes, and tsc --noEmit clean across the repo — which for this PR is the claim itself compiling. The verdict stands.