fix(context+memory+prompt): cut Jarvis multi-session context-loss across four fronts

[TYRMars]

Why

Latest persisted session 19aa5f9c-...json showed Jarvis re-running byte-identical workspace.context / todo.list / requirement.list / triage.scan_candidates calls in adjacent turns just seconds apart, ignoring the results already in context. Diagnosis (full report at ~/.claude/plans/jarvis-session-tingly-backus.md) traced four compounding causes; this PR lands fixes for all of them in one commit.

What changed

Tier	Change	File(s)
T1-A	Soften CODING_SYSTEM_PROMPT from "Before editing, call workspace.context" / "(1) call workspace.context" to conditional ("if you don't already have it from earlier in this conversation"); lead with an explicit "Reuse what you already gathered" rule.	apps/jarvis/src/serve.rs
T2-C	Strengthen DEFAULT_SUMMARY_PROMPT with BAD/GOOD examples + add strip_summary_preamble post-process to drop "The user wants me to summarise..." openers (visible in __memory__.summary*.json cache) before persisting — bails out conservatively when no paragraph break exists in the budget so a borderline first sentence is never amputated.	crates/harness-memory/src/summarizing.rs
T2-D	Replace Agent::ensure_system_prompt's insert-if-missing logic with a content-hash compare; add AgentConfig::refresh_system_prompt_on_resume (default true) so resumed conversations stop being locked into whatever prompt was active at creation time.	crates/harness-core/src/agent.rs
T3-E	Lower auto-loaded project-context cap from 32 KiB → 8 KiB so AGENTS.md/CLAUDE.md don't drown out mid-conversation tool results; truncation logs a startup WARN naming JARVIS_PROJECT_CONTEXT_BYTES so operators can opt back to a higher cap deliberately.	apps/jarvis/src/serve.rs, CLAUDE.md
T3-F	Add opt-in JsonAwareEstimator (chars * 2 / 7 + 8 for Message::Tool, falls through to CharRatio elsewhere) to compensate for the ~15-25% underestimate the chars/4 + 4 heuristic gives on dense JSON tool output. Not wired into default_estimator() to avoid silently shifting budgets in tests / fallback paths; production providers still use their TiktokenEstimator overrides.	crates/harness-core/src/memory.rs, crates/harness-core/src/lib.rs

Skipped from the original plan: T1-B (project-context hash-incremental injection). The Plan agent's premise — "every turn invalidates prefix cache" — turned out to depend on render_project_block producing different bytes turn-to-turn, which it doesn't unless project_memory files change. T1-A + T3-E together address the underlying "system prompt drowns out tool results" symptom without the inject/strip refactor risk.

Reviewer notes

• Backward compatibility: refresh_system_prompt_on_resume = true is the new default — old conversations whose persisted system prompt drifted from the binary's template will get replaced on first re-open. Operators who deliberately persisted custom per-conversation prompts can opt out via the builder.
• Wire schema unchanged: no Conversation / Message JSON fields added. Old persisted sessions load as-is.
• No env var renames: JARVIS_PROJECT_CONTEXT_BYTES continues to override; only its default lowered.
• Memory mode users: JsonAwareEstimator is opt-in only — set up via with_estimator(Arc::new(JsonAwareEstimator::new())) if you don't have a tokeniser-backed provider estimator.

Test plan

• cargo clippy --workspace --all-targets --exclude jarvis-desktop -- -D warnings — clean
• cargo test --workspace --exclude jarvis-desktop — 1100+ tests, 0 failures, including 13 new ones added here covering: summary preamble strip (5), ensure_system_prompt matrix (5), JsonAwareEstimator Tool/non-Tool/text (3)
• Manual repro (recommended): jarvis serve locally, replay the two-turn sequence from 19aa5f9c-...json — ask "项目有哪些要做的事情" then "我想优化 Web UI 体验, TODO board 是否合理" — verify turn 3 references the previous tool results instead of re-running them

🤖 Generated with Claude Code