tinymux/docs/MILESTONES.md

Milestones: Strategic Roadmap Execution

Updated: 2026-03-25 Tracks: docs/STRATEGIC_ROADMAP.md

This file is the working checklist. Each milestone has concrete exit criteria. Check items off as they land; update dates when they do.

---

Phase 1: Parity Closure (Near-term)

Unblock all remaining Tier 2 functions. Each must produce identical output to the server for its full input domain.

M1. SECURE / SQUISH — switch to co_* wrappers

co_transform (for SECURE) and co_compress (for SQUISH) already exist in the blob as Ragel-compiled, ANSI-color-aware functions. Write co_secure_wrap and co_squish_wrap, reroute s_tier2_map[], add to s_allowlist[], smoke-test.

• co_secure_wrap in softlib.c (calls co_transform with SECURE char set)
• SQUISH rerouted to existing co_compress_wrap (no new wrapper needed)
• Update s_tier2_map[] entries
• Add SECURE, SQUISH to s_allowlist[]
• Smoke clean (666 tests) — 2026-03-25
• Remove dead rv64_secure, rv64_squish

M2. TRANSLATE — ECALL or co_* wrapper

Server's translate_string() handles control characters and preserves ANSI color. Evaluate whether co_transform covers this or whether a new ECALL_TRANSLATE calling the native translate_string() is cleaner.

• Determined approach: ECALL to native translate_string() — the function uses MUX color state machine, MU_Substitutes tables, and percent-encoding logic that don't belong in the blob.
• Note: the old rv64_translate implemented 3-arg char mapping (tr() semantics), not MUX translate() semantics. Fixed to match the actual 2-arg translate(string, type) interface.
• Added ECALL_TRANSLATE (0x162) with ecall3 helper
• Add TRANSLATE to s_allowlist[]
• Smoke clean (666 tests) — 2026-03-25
• Old rv64_translate (wrong semantics) replaced

M3. STRMATCH / MATCH / GRAB / GRABALL — Unicode case fold

Server uses mux_strlwr() (Unicode case folding) before wildcard matching. The rv64 versions use ASCII tolower. Two options: (a) ECALL for mux_strlwr and keep wildcard matching in guest code, or (b) ECALL the entire match operation.

• Approach: ECALL_QUICK_WILD (0x163) calls host quick_wild() which uses mux_strlwr() for Unicode case folding. List tokenization stays in guest code; only the pattern match crosses the ECALL boundary.
• Implement (4 functions, removed wc_lower/wild_match from blob)
• Add STRMATCH, MATCH, GRAB, GRABALL to s_allowlist[]
• Smoke clean (666 tests) — 2026-03-25

M4. SORT — DUCET collation

Server supports sort types u/U/c/C using mux_collate_sortkey() (UCA-based binary sort keys). co_sort_wrap exists but needs to call the collation key generator for Unicode sort types. ASCII/numeric sort types (a/i/n/d/f) are likely fine as-is.

• Approach: ECALL_SORT (0x164) calls native sort_to_buffer() which uses the full do_asort_start/do_asort_finish pipeline with all sort types including DUCET collation (u/U/c/C), float, numeric, dbref, auto-detect. No sort logic on the RISC-V side.
• Fixed arg order in co_sort_wrap (was delim/osep/type, now matches MUX convention: sort_type/delim/osep)
• Add SORT to s_allowlist[]
• Smoke clean (666 tests) — 2026-03-25
• Add targeted sort-type smoke tests (especially u and c) — M5

M5. Parity test suite

The smoke suite (666 tests) is a baseline but doesn't specifically stress Unicode and ANSI color edge cases in the newly unblocked functions. Build targeted tests.

• testcases/tier2_parity_fn.mux — 8 targeted tests:
  • TC001: squish() ANSI color preservation
  • TC002: translate() p-mode control char conversion
  • TC003: strmatch() case-insensitive matching
  • TC004: grab/graball case-insensitive wildcards
  • TC005: sort() numeric/float/case-insensitive types
  • TC006: sort() DUCET case-insensitive (c type)
  • TC007: sort() auto-detect
  • TC008: match() case-insensitive position
• All 674 smoke tests passing with --enable-jit — 2026-03-25
• Existing tests: secure (4 tests with color+Unicode), sort (4 tests with DUCET/fullwidth), jit_parity (7 tests with chr/ord grapheme)

---

Phase 2a: ARM DBT (Medium-term)

Port ~/slow-32/tools/dbt (~23K lines) into the TinyMUX framework.

M6. Adapt ARM translator to TinyMUX DBT harness

The existing ARM DBT has its own harness (dbt.c, block_cache.c, cpu_state.h). TinyMUX has its own (dbt.cpp, dbt_state_t). Adapt the ARM translator to work within TinyMUX's dispatch loop, ECALL mechanism, and memory model.

• Map ARM DBT structures to TinyMUX dbt_state_t
• Integrate ARM emit_x64.c with TinyMUX's code cache
• ECALL dispatch from ARM-translated code
• Basic execution: simple Tier 2 function runs on ARM host

M7. ARM intrinsic registration

Register the same intrinsics (co_*, memcpy, slen, scopy, etc.) for the ARM backend so Tier 2 functions get native host calls.

• Port pretranslate_tier2() intrinsic registration to ARM
• Validate co_* functions callable from ARM-translated code
• Math intrinsics (libm) working

M8. ARM smoke tests

Full parity with x86-64: all smoke tests pass on ARM hardware.

• Cross-compile or native-build on ARM target
• 666+ smoke tests passing
• Tier 2 blob executing correctly
• JIT-compiled softcode executing correctly

---

Phase 3: JIT-by-Default (1-2 years)

Gate: Phase 1 parity closure complete. All blocked functions unblocked and smoke-tested.

M9. fun_* coverage audit

Catalog every fun_* function in the engine. For each, determine: is it reachable when JIT is on? If yes, can it be replaced? If no, it's dead code.

• Generate complete fun_* inventory (from builtin_functions[])
• Classify: JIT-handled / ECALL-handled / AST-only (FN_NOEVAL) / dead
• Document in docs/FUN_COVERAGE.md

M10. Remove --enable-jit gate

JIT becomes unconditionally on. The configure flag and all #ifdef TINYMUX_JIT guards are removed.

• Remove --enable-jit from configure.ac
• Remove #ifdef TINYMUX_JIT conditionals
• All builds include JIT
• Smoke clean on both x86-64 and ARM

M11. Remove dead fun_* code

Functions classified as dead in M9 are removed from the engine. The JIT compiler retains the semantic knowledge of what they did.

• Remove dead fun_* implementations
• Remove dead entries from builtin_functions[]
• Smoke clean
• Verify no regression in softcode behavior

---

Phase 4: Concurrency (Long-term / Research)

No milestones yet — requires design documents first. See docs/STRATEGIC_ROADMAP.md Phase 4 for the problem statement.

• Design document: mutable engine state inventory and ownership
• Design document: SQLite write coordination strategy
• Design document: scope of concurrent evaluation (what can run in parallel)