feat: allow admins to override the chat compaction model - #27151

Adds a deployment-wide 'compaction' context to the chat model override
framework. When set, chatd generates compaction summaries with the
override model instead of the chat model. Thresholds use the stricter
of the two context limits, the prompt is re-sanitized for the override
provider, and metrics plus debug runs record the model that actually
generated the summary.

Resolving the compaction model override previously built the route and
model client on every generation prepare, so a configured but
temporarily broken override could fail ordinary turns that never
compact. Split the resolver: the stored config (whose context limit
feeds the compaction trigger) is resolved on every prepare with soft
fallback, while route and client construction happen only once
compaction is known to run.

Adding the compaction context to AllChatModelOverrideContexts extends
the invalid-context error detail, so the UnknownContextReturns400
expectations must list it.

compactionNeeded at prepare time does not guarantee compaction runs:
decideGenerationAction can return FinishTurn first (completed step,
stop-after tool, max steps), and building the override client during
prepare failed such turns on hard construction errors. Carry the
resolved override config through generationCompaction and build the
client, swap identity, and sanitize the prompt inside
generateCompaction, which only runs for the compact action.

# Conflicts:
#	coderd/x/chatd/ARCHITECTURE.md

…ction

Align the cross-provider compaction prompt pass with coder/mux's
flattenProviderExecutedToolParts: rewrite provider-executed tool calls
and results into plain text parts instead of dropping them, so the
summary keeps server-side tool content (for example web_search results)
while still shedding the provider-specific wire shape other providers
reject. Anomalous provider-executed parts outside assistant messages
are still dropped.

The compaction override resolver folds the admin-configured reasoning
effort into the override config's call options, but the summary call
never consumed them. Convert the override's provider options and
resolved reasoning effort when building the override client and pass
them on the summary fantasy.Call, so a compaction model's configured
effort actually applies.

…anitize.go

The file never defined the compaction prompt; the summary prompt text
lives in chatloop/compaction.go. It sanitizes the already-built prompt
messages for the compaction override model, so name it after its entry
point sanitizeCompactionPrompt and the chatsanitize package it wraps.

The compaction trigger uses the stricter of the chat and override
models' context limits, but the still-over-limit check compared
against the chat model's limit. A stricter override then looped
through repeated compactions instead of failing the turn.