test(remote): cover the phone projection on a REAL v3 workspace - #1065

#1031 item 4.

Every phone-projection test built a v2 document by hand — `tabs[]` with a
`root` tile tree — so nothing in the suite exercised the shape the app has
actually been persisting since #1013 migrated to the unified stage. The
projection DOES handle v3, but a regression in that half would have been
silent, and the phone's entire session list depends on it.

The proof that this was a real blind spot rather than a tidiness exercise:
deleting the v3 project parsing from `agentActivity/workspaceProjection.ts`
leaves the existing phone suite completely GREEN (4 passed) and fails only the
new file. Ignoring the row's own `projectId` — the v3 membership rule — does
the same.

The input is the real `~/.config/agent-code/workspace.json` the app wrote on
2026-09-20: one window, three projects, 13 sessions across five provider
kinds. Sanitized under the same contract as the v2 fixture — structure and
every value unchanged except `cwd`, `title` and `projectTabTitle`, each mapped
one-to-one so equal values stay equal — and verified afterwards to contain no
absolute home path, username or real directory name.

Each test guards against being vacuous: the first asserts the fixture really
is v3 with no `tabs` to fall back on, the membership test asserts the file
actually places sessions in more than one project, and the identity test
asserts four or more provider kinds are represented.

The v2 test file is deliberately left alone rather than parameterised: main
reads whatever is on disk, and a window that has not saved since the upgrade
still holds a v2 document. Both generations must keep working, so both keep
their own evidence.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Review of #1065. The most important finding is that my own justification was
false, so it is corrected rather than softened.

`agentActivity/workspaceProjection.test.ts` already covers the SHARED decoder
on v3 documents, and deleting the v3 `projects` parsing fails two of its tests.
I ran that mutation only against `src/main/remote/`, saw it green, and
generalised it into "a regression would have been silent". It would not have
been. The real gap is narrower and specific to the phone: nothing drove
`RemoteWorkspaceProjection` with a v3 document at all. The header, the README
and the PR now say that instead.

The identity test asserted only `kind` and `cwd`, both pass-throughs. It now
asserts `tldrIdentity` — the key the remote server joins every TLDR and Goal
frame on, so dropping it silently empties both on the phone while the session
list still looks right. Nulling it now fails 2. The re-projection test changes
a `tldrIdentity` rather than a `title` for the same reason: that equality check
was unpinned, and removing it left all 74 remote tests green. It now fails 1.
The kind count said "four or more" under a comment naming five; it says five.

The sanitization table was wrong in three ways: it listed `projectTabTitle`,
which is v2-only and appears in neither file; it omitted `projects[].title`,
which carries three of the placeholders; and it said nothing about `drafts`,
which persists RAW COMPOSER TEXT. This capture contained none, so the file is
clean — but clean by timing, not by process, and the README now says to drop
`drafts` unconditionally when re-capturing.

`/fixture/project-N` named directories, not projects — two of them are
different directories inside one project, so the numbering never tracked
`Title N`. They are `/fixture/dir-N`.

The store fake was duplicated verbatim across both projection test files; one
definition of "the store notifies only after bytes reach disk" now lives in
`workspaceProjection.testSupport.ts`.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>