install.ps1 installs the Codex plugin and install.sh does not

[MongLong0214]

Found while deleting the dead host block for #691, and it is not caused by that
deletion — the deletion revealed it.

install.ps1:1061   & $dest plugin install-codex        ← reached
install.sh         wire_codex_plugin { … install-codex }
                   called only from wire_codex, which nothing calls

On main today, install.sh contains the Codex plugin step and never runs it.
install.ps1 runs it. The two platform installers have been doing different
things, and the shell's copy looked like evidence that they were not.

test/agent-configs.test.ts asserted the string is present in both files, which
passed — presence, not reachability. The same shape as #689: an assertion that
could not fail.

What this means for a user

A Windows install wires the Codex plugin. A macOS or Linux install does not,
unless something else does it. Whether that gap matters depends on what the
Codex plugin provides that the MCP registration does not — the shell does
write Codex's MCP config, so this is about the plugin layer specifically.

Not fixed here

#691 removes the unreachable code rather than making it reachable, because
making it reachable is a behaviour change to the shell installer and deserves its
own decision: either the shell should install the plugin (and be tested for it),
or the Windows installer should stop, and the two should be said to agree
deliberately rather than by accident.

Recording it rather than choosing, since the choice needs to know what the Codex
plugin does for a user that its MCP registration alone does not.

[MongLong0214]

The host-by-host map, so the consolidation is not guesswork

install.ps1 both delegates and wires. Line 817 calls installer-hosts; from
line ~900 it does its own work. The live call sites, resolved to hosts:

:907   generic MCP JSON writer (Wire-McpServersJson)
:967   claude-code   installed the commitlore plugin (marketplace)
:1024  codex         registered commitlore with codex mcp add
:1045  codex         config-file branch
:1065  codex         installed the commitlore plugin (marketplace)
:1107  hermes        configured MCP + skill bundle
:1129  opencode      created the config
:1142  opencode      added into an existing config

Against the enumeration in src/commands/installer-hosts.ts:

host	enumeration	install.ps1	equivalent?
codex MCP	cliHost / tomlHost	:1024, :1045	duplicated
codex plugin	added in #698	:1065	duplicated after #698
claude-code plugin	added in #690	:967	duplicated
gemini / cursor / windsurf	jsonHost table	Wire-McpServersJson	duplicated
opencode	jsonHost (json-mcp)	:1129, :1142	duplicated
hermes	delegates to hermes install	:1107	duplicated

Every row is done twice on Windows today.

Why this is not a mechanical delete

The Codex plugin gap is the proof. Before #698 the enumeration did not do
what install.ps1:1065 does — so the two sides were not equivalent, and the
shell's dead copy hid that for as long as it existed. Deleting the PowerShell on
the assumption of equivalence would have removed a step nothing else performed,
on a platform the author cannot run.

So each row needs the same treatment #698 gave the Codex plugin: confirm the
enumeration performs it, then remove the duplicate — not the reverse order.

Order

1. Confirm each row above, one at a time, against the enumeration's actual
   behaviour rather than its intent.
2. Remove install.ps1's wiring once every row is covered.
3. Then install.sh's 291 dead lines (install.sh carries seven has_/wire_ host pairs that nothing calls #691) — they become safe to delete only
   when T-1121's parity is no longer holding anything up, which is the same
   moment.

T-1121 stays untouched throughout. It is what caught this: the parity it
asserts was holding between a live script and a dead block, which is why #689 and
this issue were both invisible.

[MongLong0214]

Closed on the user-facing half — the rest moved to #691

#698 put the Codex plugin step in the enumeration, so every platform gets it.
That was the defect a user could meet, and it is fixed.

The divergence itself is not. install.ps1 still delegates at line 817 and
wires hosts itself from line 1089, so Windows does every host twice. The
host-by-host map is in the comment above and the work is tracked on #691, which
stays open for exactly this reason — the two are one change:

1. confirm the enumeration performs each row
2. remove install.ps1's duplicate
3. then install.sh's 291 dead lines

The order is not a preference. The Codex gap proved the two sides are not
equivalent, so removing a duplicate before confirming its replacement deletes a
step nothing else performs — on a platform the author cannot run.

Reopening this issue is not needed; #691 carries it. Recorded here so the closure
is not read as "the installers agree now".