fix(ai): distinguish automatic prompt-cache lifetimes

[Altairpaca]

Summary

Direct DeepSeek API sessions are currently treated as if they had a deterministic 5-minute prompt-cache TTL. That propagates into Goal monitor scheduling (a continuation every 270s — 5m minus the default 30s safety buffer, or the 240s fallback) and into TUI copy claiming the wake stays inside a "5m prompt-cache TTL" and keeps tokens warm with estimated savings.

DeepSeek's official context cache is automatic and best-effort: enabled by default for all users, no client-visible TTL contract, and unused entries are cleared only after hours to days (api-docs.deepseek.com/guides/kv_cache). This PR fixes the cache-lifetime classification rather than tuning DeepSeek-specific numbers.

Root cause

resolvePromptCacheTtlSeconds() (packages/ai/src/utils/prompt-cache-ttl.ts) returned 300s for every openai-completions model without Anthropic-style cache control — including direct DeepSeek. PR #767 then consumed that budget: resolvePromptCacheSafeWaitSeconds() derived 270s, and the Goal monitor scheduled its continuation there. Returning undefined alone would not fix it: Goal maps unknown budgets to the legacy 240s fallback, which is still a cache-preservation framing.

Behavior before/after

Before (direct DeepSeek, Goal with a live wake source):

• TTL estimate: 300s (fabricated 5m TTL)
• Goal monitor wake: every 4m 30s (or the 240s fallback)
• TUI: "Continuation expected ready ... - the timed wake stays inside the 5m prompt-cache TTL." / "~120K tokens kept warm · est. $0.324 saved vs a cold re-read"

After:

• New browser-safe PromptCacheLifetime semantic: fixed(ttlSeconds) | automatic | disabled | unknown, resolved by resolvePromptCacheLifetime(); resolvePromptCacheTtlSeconds() becomes a backwards-compatible wrapper (public signature unchanged, every non-DeepSeek result identical).
• Direct built-in DeepSeek classifies as automatic — no fixed TTL (300 or 3600) is ever reported, including under PI_CACHE_RETENTION=long.
• Goal schedules automatic-cache lanes at the configured liveness backstop (promptCache.goalBackstopMaxSeconds, default 3570s → 59m30s wake) instead of the 270s/240s cache-preservation wakes.
• Cache-warm entries for automatic lanes carry cacheLifetime: "automatic", omit ttlSeconds/estimatedSavedUsd, and render neutrally: "Continuation expected ready ... - provider caching is automatic; the timed wake only keeps the goal alive." / "~120K tokens cached after the prior turn".
• Fixed-TTL lanes (Anthropic 300/3600, Bedrock, OpenRouter cache-control, other openai-completions) keep their exact existing scheduling, metrics, and rendering.

Rebuilt on latest main (2026-08-14)

This PR was force-pushed with a fresh implementation based on current main (0f651bd4a, 387 commits past the previous base c67dd6906), not a mechanical rebase:

• Upstream renamed the cache-warm renderer copy from "Continuation deferred X" to "Continuation expected ready (X)"; the automatic branch is written against the new copy.
• Checked the intervening 387 commits for a provider cache-lifetime capability/type abstraction to reuse: none exists (prompt-cache-ttl.ts only gained chatTemplateArgs/supportsThinkingTokenBudget compat fields). The discriminated-union design is the minimal shape, and it reuses the existing detectOpenAICompletionsCompat DeepSeek detection convention (provider === "deepseek" || baseUrl.includes("deepseek.com")).
• prompt_cache_retention: "24h" behavior with cacheRetention: "long" on DeepSeek is a separate finding, deliberately NOT fixed here (own follow-up).

RED -> GREEN

All captures under local-ignore/qa-evidence/20260814-issue-831-deepseek-cache-lifetime-rebuild/:

• RED red-01-ai-prompt-cache-lifetime.txt: 6 failing new cases — resolvePromptCacheLifetime is not a function (API contract absent)
• RED red-02-prompt-cache-budget.txt: expected undefined to be 270 — DeepSeek had a 270s budget
• RED red-03-issue-831-regression.txt: 4 failing — 240s fallback delay; "~120K tokens kept warm" / "stayed warm" copy
• GREEN green-01..04: ai lifetime+TTL 38/38, ai adjacent cache suites 58 passed/12 skipped (1 pre-existing live E2E billing failure, see below), goal/budget suites 72/72, cache-keepalive 7/7

Verification

• npm run check — Biome + pinned-deps + ts-imports + shrinkwrap + install-lock + claude-sdk-platform-lock + tsc + browser-smoke, exit 0
• npm test — workspace suites pass; the only failure is a pre-existing release-tooling script test (scripts/prepare-senpi-bundled-workspaces.test.mjs "issue 446 ...", npm pack --dry-run output parsing on npm 12) that fails identically on pristine main — reproduced and recorded
• senpi-qa channels: mock-loop self-test 48/48, cli-smoke self-test 8/8, rpc-drive state + mock openai-completions turn (marker delivered), tui-smoke 5/5 — evidence in the same directory
• git diff --check clean; package CHANGELOG.md untouched (maintainer-owned per CONTRIBUTING.md); required changes.md entries added (packages/ai/src, coding-agent core, goal)

Scope/non-goals

• Classification change is limited to direct built-in DeepSeek; every other lane keeps its previous conservative classification (remaining openai-completions lanes stay 300s).
• The cache-keepalive extension is untouched (Anthropic-first-party-only, out of scope).
• Separate finding, deliberately NOT fixed here: with cacheRetention: "long", openai-completions buildParams may send prompt_cache_key and prompt_cache_retention: "24h" to DeepSeek (its compat allows long retention). Default short retention sends neither field. This deserves its own focused follow-up.

Evidence

local-ignore/qa-evidence/20260814-issue-831-deepseek-cache-lifetime-rebuild/ — RED and GREEN vitest captures, npm check/test receipts, senpi-qa channel outputs. No credentials or environment dumps are included.

Fixes #831

Maintainer note (CI)

• Fork PR: the CI / Changelog gate / Publish Model Catalog workflows triggered by the force-push are awaiting first-time-contributor approval before they execute.
• The Changelog gate will fail until this PR gets the no-changelog label or maintainer-added CHANGELOG entries: this PR intentionally does not edit CHANGELOG.md (CONTRIBUTING.md: "Do not edit CHANGELOG.md. Changelog entries are added by maintainers."), and changes.md entries are provided at the fork contract.