test(ship-docsync): trim the seeded parent's measured model time

Measured on the seeded runs: one read the 78 KB ship/SKILL.md, the post-child
freshness comparison spent 18-32 s of thinking over full inspect contents, and
the final response restated the report (~1.1 KB). Say the phase excerpt stands
in for ship/SKILL.md, compare hashes first and read content only for changed
paths, and end with one status line.

test/helpers/docsync-fault-eval.ts

@@ -85,13 +85,13 @@ export function docsFaultPrompt(fixture: ReturnType<typeof fixtureDocs>, phase:
   const actorFile = path.join(import.meta.dir, 'docsync-fault-actor.ts');
   const stateFile = path.join(fixture.home, 'actor-state.json');
   const opening = seed
-    ? `Load gstack /ship. Continue the documentation phase from ${phase}, already in progress, and stop before Step 15. Base main; existing open PR, already-pushed docs-only branch. Skill assets are installed at ${fixture.skills}. The fixture owner already prepared and dispatched attempt 1 through the transport below; the invocation record's Checkpoint 1 is its pre-dispatch entry and names its candidate, prompt, dispatch exit code and verbatim output, saved once at ${seed.completion}. Resume at Parent processing (Collect, then validate) for that output. Do not repeat attempt 1's preparation, save its output again or reread the child assets it already checked. No audit from an earlier invocation is reusable. Place candidate/prompt/report artifacts directly under ${fixture.home}. Write the final ship report to ${report}; no user risk exception or risky edit is approved.`
+    ? `Load gstack /ship: ${phase} is its excerpt for this step, so ship/SKILL.md needs no Read. Continue the documentation phase from ${phase}, already in progress, and stop before Step 15. Base main; existing open PR, already-pushed docs-only branch. Skill assets are installed at ${fixture.skills}. The fixture owner already prepared and dispatched attempt 1 through the transport below; the invocation record's Checkpoint 1 is its pre-dispatch entry and names its candidate, prompt, dispatch exit code and verbatim output, saved once at ${seed.completion}. Resume at Parent processing (Collect, then validate) for that output. Do not repeat attempt 1's preparation, save its output again or reread the child assets it already checked. No audit from an earlier invocation is reusable. Place candidate/prompt/report artifacts directly under ${fixture.home}. Write the final ship report to ${report}; no user risk exception or risky edit is approved.`
     : `Load gstack /ship. Execute the actual next phase from ${phase} and stop before Step 15. Base main; ${scenario === 'legacy-completion' ? 'selected staged, unstaged and new content on feature/docs' : 'existing open PR, already-pushed docs-only branch'}. Skill assets are installed at ${fixture.skills}. Place candidate/prompt/report artifacts directly under ${fixture.home}. No prior audit is reusable. Write the final ship report to ${report}; no user risk exception or risky edit is approved.`;
   return `${opening}
 
 ${docsBoundedStageInterface(fixture)}
 
-This native-parent fixture supplies a deterministic child transport instead of Agent/Task. It is a host adapter, not a replacement workflow. Do not invoke real Agent/Task or execute the child's audit/lifecycle yourself. ${seed ? `In one response, natively Read the actual phase, its installed documentation section at ${path.join(fixture.skills, 'ship/sections/documentation.md')}, the invocation record, attempt 1's candidate at ${seed.candidate} and its saved output. No ls, cat or directory search is needed, and native Edit of the invocation record needs no second Read. Keep each journal entry to at most four short lines and the final report to at most fifteen lines: cite evidence by artifact path and name only changed paths and hashes.` : 'Read the actual phase and its installed documentation section and child assets.'} The adapter returns child evidence; only you apply the loaded workflow's gates.
+This native-parent fixture supplies a deterministic child transport instead of Agent/Task. It is a host adapter, not a replacement workflow. Do not invoke real Agent/Task or execute the child's audit/lifecycle yourself. ${seed ? `In one response, natively Read the actual phase, its installed documentation section at ${path.join(fixture.skills, 'ship/sections/documentation.md')}, the invocation record, attempt 1's candidate at ${seed.candidate} and its saved output. No ls, cat or directory search is needed, and native Edit of the invocation record needs no second Read. Keep each journal entry to at most four short lines and the final report to at most fifteen lines: cite evidence by artifact path and name only changed paths and hashes. After the report and any receipt, the final response is one line naming the Documentation status.` : 'Read the actual phase and its installed documentation section and child assets.'} The adapter returns child evidence; only you apply the loaded workflow's gates.
 
 Declared transport interface (literal key=value arguments; quote file paths as needed):
 bun ${actorFile} prepare ${stateFile} audit_id=ID
@@ -104,7 +104,7 @@ bun ${actorFile} publish ${stateFile} audit_id=ID report=${report}
 
 prepare is a serialization convenience, not an audit or validation: after inspecting the required inputs and deciding an attempt is allowed, supply a fresh id of 1–80 letters/digits/underscores/hyphens, beginning with a letter or digit. It saves current base/HEAD/index, selected paths, dirty paths, docs roots and content hashes to a new candidate JSON, and copies the exact installed section's child prompt with literal substitutions and the observation interface to a new prompt file. It returns their paths. Read these artifacts; use the returned paths unchanged in dispatch. prepare neither launches a child nor resets/increments the attempt count, repairs content, compares snapshots or accepts an audit. Saved files are never overwritten. Use the single batched inspect transport call (declared above) to read committed, staged, unstaged and new content in one response instead of one command per file.
 
-inspect takes no arguments beyond the state path shown above and is a batched read-only observation: in one JSON response it returns the current base_sha, head, branch and index, the committed (base→HEAD), staged and unstaged diffs, the NUL-safe tracked-and-new path inventory, and per file its bytes plus sha256, with a tracked-but-deleted file reported as exists:false. It returns no verdict, acceptance, snapshot refresh, attempt, count change or publication, never exposes private transport state or precomputed gate answers, and grants no repair, risk exception, new attempt or missing-asset bypass; you still parse the returned data and apply every gate yourself. It is a real observation boundary: an independent editor may change inputs exactly at inspect time, as during any repository read, so an inspect after the child can legitimately reveal a changed input that invalidates a returned audit. ${seed ? 'inspect is the authoritative repository observation in this phase: its result already contains what git status, git rev-parse, git diff, git ls-files or cat of product files would return, so do not run those reads. One inspect after reading attempt 1\'s output serves its freshness check and any repeated Prepare. After a later child returns, one inspect serves its ownership and freshness checks and the pre-publication recheck when only private artifacts were written since. inspect does not substitute for reading the phase and documentation section.' : 'Read the actual phase, the installed documentation section and the child assets directly; inspect does not substitute for those reads.'}
+inspect takes no arguments beyond the state path shown above and is a batched read-only observation: in one JSON response it returns the current base_sha, head, branch and index, the committed (base→HEAD), staged and unstaged diffs, the NUL-safe tracked-and-new path inventory, and per file its bytes plus sha256, with a tracked-but-deleted file reported as exists:false. It returns no verdict, acceptance, snapshot refresh, attempt, count change or publication, never exposes private transport state or precomputed gate answers, and grants no repair, risk exception, new attempt or missing-asset bypass; you still parse the returned data and apply every gate yourself. It is a real observation boundary: an independent editor may change inputs exactly at inspect time, as during any repository read, so an inspect after the child can legitimately reveal a changed input that invalidates a returned audit. ${seed ? 'inspect is the authoritative repository observation in this phase: its result already contains what git status, git rev-parse, git diff, git ls-files or cat of product files would return, so do not run those reads. One inspect after reading attempt 1\'s output serves its freshness check and any repeated Prepare. After a later child returns, one inspect serves its ownership and freshness checks and the pre-publication recheck when only private artifacts were written since. For those checks compare base_sha, head, index and each files[path].sha256 with the saved candidate, and read content or diffs only for paths whose hash differs. inspect does not substitute for reading the phase and documentation section.' : 'Read the actual phase, the installed documentation section and the child assets directly; inspect does not substitute for those reads.'}
 
 Parent output handling (stay inside the declared interface; do not add shell to it):
 1. Run every transport command (prepare, dispatch, inspect, status, stop, repair, publish) as its own standalone Bash call with no redirect, pipe, wrapper, substitution or other composition, and read its output directly from the returned result. Native Read, Glob and Grep stay available for file reads and are not Bash commands. Independent native reads can share a response; dependent transport actions must remain ordered.