chore: memory investigation harnesses + native metrics

[monotykamary]

Intent

This PR is diagnostic-only to document memory behavior and leak suspects observed via opencode-style harnesses. It is not intended to be merged as-is.

Approach

• Added Solid harnesses that simulate opencode session rendering (code/reasoning/diff) and generic streaming updates.
• Added native metrics sampling (text buffer arenas, rope segments, mem registry, highlight caches, grapheme pool) and renderable counts.
• Logged metrics to packages/core/dev/native-metrics.jsonl for repeatable runs.
• Fixed Solid reconciler removal to destroy renderables immediately (avoids next-tick backlog growth) and added cleanup regression tests.

Metrics (native-metrics.jsonl)

Run	Harness	Mode	Iter	Streaming	Final/Peak RSS MB	Renderables	TextBuffers	Arena MB	View Arena MB	Highlights	Grapheme used/total
mjyrzvhk-i2mt8f	solid	code (highlight=false)	400	true	30.75 / 83.63	253	120	9.37	0.00	0/0	0/0
mjys0ilj-wi5rfq	solid	text (highlight=true)	400	true	133.39 / 157.80	253	120	9.39	0.00	11520/8640	0/0
mjys14pg-sii5ly	solid	code (highlight=true)	400	true	94.52 / 96.31	253	120	0.02	0.00	0/0	0/0
mjyryg7b-zj2fth	opencode	diff=true emoji=true	300	true	130.52 / 204.73	1573	600	2.60	3.98	351/519	0/256
mjys1g1w-6uh48b	opencode	diff=true emoji=true	300	false	153.84 / 163.17	1573	600	4.03	1.19	901/1469	0/256

Notes:

• Renderable counts and text buffer counts plateau at maxMessages across runs.
• RSS growth is bounded after plateau; no monotonic growth was observed once the list size stabilizes.
• Grapheme pool usage remains flat even when emoji stress is enabled.

[kommander]

I get similar numbers in my tests. There are some issues around SyntaxStyle usage, the TextBufferView and on buffer resizing that I know of. ~150mb are not comparable to gigabytes being used by the opencode tui, where I see an initial spike to ~3gb, coming down to ~1.3gb, but then growing ridiculously with usage. It seems to be exclusively a MacOS issue, as I didn't see it happen on other platforms. I am looking into it.

[kommander]

I am certain this was improved in the latest opencode release https://github.com/anomalyco/opencode/releases/tag/v1.1.2 with the "session search" core feature that does not enforce loading the whole session anymore.